Premium A+ Content is now open to Brand Registered sellers in Seller Central without the old barriers. No invite, no waitlist, no minimum spend, no requirement to have published a stack of standard A+ projects first. The modules that used to sit behind a vendor relationship or a big budget, video, hover hotspots, interactive comparison tables, are available to a two-person brand selling one product.

The second change is the one to read carefully. Since late July 2026, sellers have had to tag images and A+ content containing photorealistic AI-generated people before uploading them.

The rule, in plain terms

If an image or video in your listing or A+ content contains a photorealistic AI-generated person, you have to declare it in the file’s metadata. Specifically, the keyword contains-synthetic-performer goes into the dc:subject (XMP) field, which you set with an IPTC-compatible metadata editor before upload. Amazon then shows a customer-facing indicator on the image.

It does not apply to images of real people, even heavily edited ones. It does not apply to non-photorealistic illustrations, cartoon characters, or images with no people in them. Forbes reported on Amazon’s requirement to label AI-generated people in listing images, which follows New York’s synthetic performer disclosure law that took effect in June.

The consequence for skipping it is not a warning email. Amazon can flag the media, and a listing whose main image is non-compliant can be suppressed from search until it is fixed. For a seller whose entire revenue runs through one ASIN, that is the difference between a normal Tuesday and no Tuesday.

Why these two changes belong in the same conversation

Premium A+ becoming free means a lot more sellers are about to build a lot more image-heavy content, quickly, on a budget. The cheapest way to fill a comparison module or a lifestyle banner is a generated model. So the exact population that just got access to the format is the population most likely to trip the disclosure rule, and least likely to have anyone checking image metadata.

That is the practical risk. Not the rule itself, which is reasonable and narrow. The gap between a free creative upgrade and a technical compliance step that lives in a file field nobody on a small team has ever opened.

The workflow that prevents the problem

The fix is a process change, not a tool purchase.

Decide the policy once, per image type. Product-only shots, no people, no issue. Lifestyle shots, decide now whether you use real photography, licensed stock of real people, or AI-generated people. Pick one per category and write it down. Ambiguity is what causes the miss.

Tag at creation, not at upload. If someone generates an image with a person in it, the metadata goes in at that moment, in the same session. Tagging at upload time means the person doing the upload has to remember which of forty files had a generated model in it. They will not.

Keep a source column in your asset list. One spreadsheet column: real photo, licensed stock, AI-generated, AI-edited real person. That single field answers the compliance question in a second and survives staff turnover.

Audit what is already live. If you have used generated people in the past year, those assets are sitting in current listings. Go back through them before Amazon does.

Where I would not rush

Free access to Premium A+ does not mean it is automatically worth building. The modules take real design work, and a badly built premium module converts worse than a clean standard one. Amazon has also introduced a quality scoring layer that looks at completeness, image quality and module mix, so a rushed job is now visible rather than just ineffective.

If you sell a simple, well-understood product where buyers decide on price and reviews, the upgrade may add very little. Premium A+ earns its keep on products that need explaining: technical specs, sizing, comparisons across a range, anything where a buyer is choosing between your three variants and cannot tell them apart.

Start with your one highest-revenue ASIN. Build it properly. Measure the conversion rate before and after. Then decide whether the catalogue is worth doing.

The wider signal for anyone not selling on Amazon

A disclosure rule that lives in file metadata rather than in a checkbox is a preview of where this is going. Platforms are moving from asking whether content is AI-generated to reading whether it says so about itself, at the file level, automatically. YouTube went down a similar path with its originality rules for AI video. Amazon is now doing it for still images.

Marketing teams that cannot answer “where did this asset come from” for every file in their library are going to keep meeting rules they cannot comply with. The provenance question is not a legal nicety anymore. It is becoming an operational requirement, and the teams that logged it from the start are going to find the next three years much cheaper than the ones who did not.

If you have generated people anywhere in your Amazon catalogue, check those files this week. The rule is already live, and the enforcement mechanism is search suppression, not a friendly nudge.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Amazon Just Made Premium A+ Free for Every Brand. The Catch Is in Your Image Files. appeared first on Mark8ng.com.

Search Engine Roundtable covered it under a headline that says the quiet part plainly: AI Overview local results are pulling from low-quality listicles. A self-promotional roundup, written by a competitor, became the authoritative voice describing several businesses to their own potential customers.

It is tempting to file this under “AI is still buggy” and move on. That reading misses what it tells you.

What this actually reveals about how the answer gets built

A business website is a first-person source. It says “we do X, we serve Y area, we have been doing this since 2011.” A listicle is a third-person source. It says “Firm A is best for personal injury, Firm B is best for family law, Firm C has the fastest response times.”

The second format is directly usable as a comparative answer. The first is not. When a model is asked to produce one short, differentiating sentence about each of five nearby businesses, the listicle already contains exactly that sentence, in exactly that shape, for all five. Your homepage contains a sentence about you, written the way you want to be described, in a format that does not slot into a comparison.

So the model does not pick the listicle because the listicle is better. It picks it because it is structurally convenient. Quality did not lose to spam. Quality lost to formatting.

That is the uncomfortable finding, and it is a different problem from the one most people are shouting about.

Where this gets messy

The obvious response is to write your own listicle. Publish “The 7 Best Plumbers in Leeds”, put yourself at the top, and hope the model quotes you describing yourself.

Plenty of people are already doing this, which is exactly why the citations look the way they do. It works for a while. It also produces a web where every business in a category has written a ranking of its own category, all of them self-serving, all of them technically citable. If that is where this ends up, the citations stop meaning anything and Google will eventually have to demote the whole format. Anyone who built their visibility on it will lose it in a single update.

There is a real accountability gap underneath this too. The businesses being described by a competitor’s article had no part in it, no notification, and no correction path. They are not misrepresented by a review they can flag or a profile they can edit. They are misrepresented by a third party’s marketing asset that the AI decided to trust. Nothing in the current system gives them standing to object. That is a harder version of the problem covered in what you can actually correct when an AI Overview gets your business wrong, because here there is no source you own to fix.

What actually matters instead

If comparative framing is what gets cited, the useful move is to publish comparative information about yourself honestly, rather than to publish a fake ranking.

That means writing the sentences a comparison needs, on your own site, in plain language. Who you are best for. Who you are not right for. What your typical turnaround is. What you charge, or at least what range. What you do not do. A page that says “we handle residential conveyancing in West Yorkshire, typically completing in eight to ten weeks, and we do not take commercial work” gives a model something usable in a comparison without inventing a ranking.

Most businesses will not do this, because saying who you are not for feels like turning away work. That is precisely why the ones who do will get quoted.

The risk is real, though. Publishing “we do not do X” can genuinely cost you enquiries from people who would have called anyway. For a very small business with low lead volume, that trade may not be worth it. This works better when you have enough enquiries to want better-qualified ones rather than more of them.

A simple version for small teams

If you have an afternoon, not a quarter:

Search your main service plus your town on a phone, logged out. Note which sources the AI cites for the businesses in the local pack. If they are directories and listicles rather than business websites, you now know the shape of the gap.

Then add one section to your main service page: a short, honest “who this is for and who it is not for” block, with specifics. Not marketing language. Actual constraints, areas, timescales, and price ranges.

Then check back in a month. Not because it will definitely have changed, but because the only way to know whether any of this works for your market is to watch your own citations, and almost nobody does.

The thing to hold onto: an AI answer is a summary of what is easiest to summarise, not a ranking of who is best. Treating it as a verdict on quality will send you chasing the wrong fix. Making yourself easy to describe accurately is a slower play than gaming the format, and it is the one that survives the next update.

If a competitor’s listicle is currently describing your business to your customers, that is worth knowing this week. The correction is not a complaint to Google. It is publishing the better version of the sentence yourself.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Google’s AI Is Quoting a Competitor’s Listicle Instead of Your Website appeared first on Mark8ng.com.

Business Insider found similar cases: a property manager in Idaho whose firm was merged in the AI summary with a similarly named company that had closed years earlier, and a consultancy labelled a scam on the strength of an old Reddit thread about an unrelated app.

What makes this different from the usual “Google got something wrong” complaint is the shape of the answer. A list of ten blue links gave a business room to exist in position four while a bad review sat in position two. One AI summary does not. It reads as a verdict, most people do not scroll past it, and there is no obvious button marked “this is about a different company.”

Where the wrong information is actually coming from

Before you can fix anything, you need to know which layer produced the error. In practice there are four, and they have completely different repair paths.

Your Business Profile. Name, category, hours, address, service area, phone. This is the layer you control most directly and the one most small businesses have half filled in.

Review text on Google itself. Reviews attached to your profile, including ones left by mistake on the wrong listing.

Third-party pages. Directories, aggregators, old listings, forum threads, review sites you have never signed up to, and roundup articles that mention you in passing. This is where most of the genuinely damaging errors originate.

Entity confusion. Google has decided your business and another business are the same thing, or that a review about a different company belongs to you. This is the hardest one, and it is what hit The Plastics Shed.

Most people jump straight to complaining about the AI. The AI is usually not inventing anything. It is faithfully summarising sources that were already wrong, and doing it with more confidence than those sources deserved.

What you can actually correct

Start with the things that have a real mechanism behind them, in this order.

1. Fill in the Business Profile properly. Not the minimum. Every field, including the description, the service list, the attributes, and the categories. A sparse profile leaves the model reaching for outside sources to answer basic questions. You can edit your Business Profile directly from Search or Maps once you are verified.

2. Report the specific AI Overview. There is a thumbs down and a “report a problem” option on the overview itself, and you can submit feedback on an AI Overview with a written explanation. Be specific. “This is wrong” does nothing. “This summary attributes reviews from a different company with a similar name to my business, my correct profile is X” gives a human reviewer something to act on.

3. Flag misattributed reviews individually. If reviews on your own listing are clearly about another company, report each one rather than reporting the listing.

4. Fix the source pages you can reach. Directory listings with an old address, an aggregator that has your business under a defunct trading name, a partner site with wrong opening hours. These are boring and they are usually the real cause. Work through the first three pages of results for your business name.

5. Publish a page that answers the question directly. If the bad summary is about your reliability, refund policy, or whether you are still trading, put a clear, plainly worded page on your own site covering exactly that. Models prefer a direct, unambiguous answer over an inference from scattered mentions.

What you cannot correct

This is the part most guides skip.

You cannot force a correction on a timeline. The cases in the reporting took weeks, and there is no service level attached to any of these reports. You cannot remove a legitimate negative review from a third-party site just because the AI weighted it oddly. You cannot stop a self-promotional roundup article from being cited instead of your own site. And you cannot get money back for ads that ran alongside a summary that was harming you, at least not on current evidence.

There is also no way to check your work reliably. AI Overviews do not fire for every query, they vary by device and location, and the summary you see is not necessarily the one your customer sees. Testing means checking the same query several times, on different devices, and ideally asking someone outside your area to try it.

The boring part that actually matters

If your business name is close to another business name in the same country, you have a structural risk here, not a one-off incident. A café called The Green Door in one city and a nightclub called The Green Door in another will get confused eventually. So will two consultancies with the same founder surname.

The defence is unglamorous: be the most complete, most consistent, most machine-readable version of yourself across every place that mentions you. Same legal name, same address format, same phone number, same category language. Consistency is what lets a model separate you from the other one. Every place you are described differently is a place the confusion can start.

Most small businesses do not have a reputation problem. They have a data consistency problem that only becomes a reputation problem once a model summarises it.

When this is not worth your time

If the incorrect summary appears on a query nobody searches, leave it. Chasing corrections on low-volume brand variants will eat days and change nothing. Check your actual search terms first. If the wrong answer only appears when someone types your business name plus a word no customer would ever use, it is noise.

Equally, if you are a small local business and the error is minor, opening hours off by an hour, an old category, fix the profile and move on. The weeks-long escalation path is for summaries that genuinely misrepresent whether you are trustworthy or still trading.

What is worth doing this week, regardless: search your own business name, on a phone, logged out, and read what the AI says about you. Most owners have never done it. Some of them are about to find out why their enquiries dropped.

This is the kind of monitoring mark8ng.ai is being built around: knowing what machines are saying about you before a customer tells you, and fixing the source instead of arguing with the output.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Google’s AI Is Describing Your Business Wrong. Here Is What You Can Actually Correct. appeared first on Mark8ng.com.

Seven BdThemes plugins were used to create hidden administrator accounts and install webshells on live sites. No plugin file was modified. No update was required. The plugin code in the WordPress.org repository was, and remained, exactly what it was supposed to be.

What actually happened

BdThemes plugins include a component that pulls promotional banners from a remote API and displays them inside the WordPress admin dashboard. An attacker compromised that upstream feed and changed what it returned. The plugin then did what it was built to do: fetch the content and render it in wp-admin.

The underlying flaw is a cross-site scripting issue rated 5.4, medium severity, apparently introduced in Prime Slider 4.1.9 on 1 March 2026 before spreading to other products in the range. On its own, a medium-severity XSS in an admin banner is the kind of finding that sits in a backlog for months. Combined with control of the feed it renders, it became an admin takeover.

The payload runs in the logged-in administrator’s own browser session. It creates a new administrator account, either through the REST API or the standard user creation flow, then downloads and installs a plugin with a plausible name, one reported example being wp-smart-thumbnails, containing a webshell for persistent access. Everything happens with your credentials, from your browser, on a site where nothing on disk looks unusual.

Wordfence was notified on 7 August. The earliest possible start of the campaign is 23 June, which is the number that should worry you, because it means “I checked last week and it was fine” covers roughly the last six weeks of a seven week window.

The affected plugins are Element Pack Addons for Elementor, Prime Slider Addons for Elementor, Pixel Gallery Addons for Elementor, Ultimate Post Kit, Ultimate Store Kit, Live Copy Paste and Smart Admin Assistant. They were pulled from the repository while the WordPress Plugins team investigates.

The twenty minute check

If any of those plugins have been on your site since June, do this today rather than this week. If you are not sure, check anyway, because Elementor add-on packs are exactly the kind of thing a previous agency installs and nobody remembers.

1. Users, then Administrator. Count them. You should be able to name every single one out loud. Any account you cannot immediately account for is the finding, regardless of how ordinary the username looks.
2. Sort users by registration date. Anything created since 23 June that you did not create yourself gets investigated before anything else.
3. Plugins list. Look for anything installed that nobody on your team installed, particularly names that sound like generic utilities. Check the plugin’s page on WordPress.org actually exists.
4. Application Passwords. Open each admin profile and look at the bottom. Any application password you did not create is a live credential that survives a password reset.
5. Users again, after cleanup. Removing a rogue admin without removing its application passwords and its webshell plugin achieves nothing.

If you find something, assume the site is compromised rather than assume you caught it early. That means restoring from a backup predating 23 June if you have one, or bringing in someone who does incident response for a living. A marketing team cleaning up a webshell by deleting the file it can see is the most common way sites get reinfected a fortnight later.

The separate thing you also need to patch

Running alongside this, and unrelated to it, WordPress shipped an emergency fix on 6 August for a flaw researchers named XSS2Shell, tracked as CVE-2026-64638 and rated 8.9. It starts as an unauthenticated cross-site scripting bug on the login screen, caused by two of WordPress’s own sanitisation routines disagreeing about what counts as a tag, and chains through to PHP execution on the server. The chain needs a logged-in administrator to click something on an attacker-controlled page, which in the researchers’ demonstration is a single ordinary click.

The fix is WordPress 7.0.3. There is no confirmed exploitation in the wild yet. If your site auto-updates core, you are probably already on it, and it takes ten seconds to confirm.

The habit that actually protects you

Here is the uncomfortable part. Updating everything on schedule would not have prevented the BdThemes compromise. The plugins were legitimate, current, and correctly signed. The malicious content arrived over the network at runtime, into the one part of your site you assume is safe because you are the only one who sees it.

Which means the control most teams treat as their security posture, keeping plugins updated, was never the control they thought it was. It is necessary and it is not sufficient. The thing that would have caught this in a day is much less impressive: knowing exactly how many administrator accounts your site is supposed to have, and looking at that list on a fixed schedule.

Make it monthly, put it in a calendar, and write the expected admin count somewhere you will see it. A site with three admins and a written note saying “three admins: Sarah, Tom, the developer” catches a fourth in about four seconds. A site where nobody knows the expected number will not catch it at all.

What to avoid

Do not respond to this by ripping out every Elementor add-on you have. Plugin count is a real risk factor, but panic-removing page builder dependencies on a live commercial site tends to break layouts on pages that make money, and the tradeoff rarely favours the person doing it at 6pm.

Do not assume a managed host has handled it. Host-level scanning looks for modified files and known malware signatures. A legitimate plugin rendering a poisoned banner, and a user account created through the official REST API, look like normal site activity from the outside.

And be careful about the opposite error. If you run a small brochure site with two plugins and one admin, this is a ten minute check and then you are done. Not every security story requires a project.

The reporting on the mechanism, including the payload behaviour, is set out in the BdThemes supply chain analysis, and the login screen flaw is documented in this WordPress pre-auth XSS write-up. We covered a more conventional version of this problem last week in the WooCommerce login plugin that handed out admin access, where updating was the fix. This time it was not.

Go and count your administrators.

Editor’s note: This is an active investigation and details may change, so check the latest vendor advisories before making decisions about your site.

The post Seven WordPress Plugins Created Hidden Admins Without Changing a Single File appeared first on Mark8ng.com.

This has a name now. Kevin Indig called it the ghost citation, and a study he ran with Semrush put a number on how common it is. Across 3,981 domain appearances from 115 prompts, run in 14 countries across ChatGPT, Gemini, Google AI Overviews and Google AI Mode, 61.7% of appearances were citations with no brand mention in the answer. Only 13.2% got both. Put differently: 74.9% of appearances included a citation, but only 38.3% included the brand name.

Your content can be doing all the work while your brand stays completely invisible to the person reading.

The engines are not doing the same job

The split between engines is severe enough that a single visibility score is close to meaningless.

• Gemini names the brand in 83.7% of appearances but generates a citation link only 21.4% of the time. It answers from what it already knows about you.
• ChatGPT does the reverse: cites 87% of the time, names the brand in 20.7%. It reads like a paper with footnotes.
• Google AI Overviews sit between the two and lean towards citing.

In the study’s data there was almost no overlap between the brands ChatGPT cited and the brands Gemini named for the same prompt. On 100 of 454 prompt and domain combinations, the engines disagreed on whether to name the brand at all. So if you are tracking one engine and reporting it as “AI visibility”, you are reporting one behaviour of one system.

The content type doing the damage is the one you produce most

This is the finding that should change what a small marketing team publishes next month.

Informational content, the “what is”, “how does”, “explained” format that most business blogs are built from, earned an 89.3% citation rate and an 18% mention rate. Comparative content, the “best”, “versus”, “recommended” format, produced a 43.3% mention rate, roughly 2.4 times more brand mentions. How-to content sat at 42.8%. Commercial queries reached 35.6% mentions against an 84.4% citation rate.

The logic is uncomfortable but obvious once you see it. Informational content is raw material. The model reads it, absorbs the fact, writes the fact in its own words, and links the source out of politeness. Comparative content forces the model to name things, because a comparison without names is not a comparison.

So the agency that spent eighteen months building a definitive glossary of industry terms has produced the single most citable and least brand-visible asset available. Meanwhile the competitor with six honest “X versus Y” pages, including the cases where they lose, gets named.

Why the model will not repeat your claim

There is a second layer under the format question, and Forbes framed it well this week: every brand claim an AI repeats without independent backing is a claim the AI is vouching for itself. Systems tuned against Google’s quality rater guidelines weight first-hand experience heavily, and a product page describing your own jacket as the warmest available has an obvious incentive problem. A forum thread describing a Colorado winter in that jacket does not.

This is where most GEO advice goes vague, so here is the concrete version. If your only evidence for a claim lives on your own domain, expect the model to use the surrounding information and skip the claim. Corroboration that actually shifts this looks like: a named customer willing to be quoted with specifics, a third-party review platform with enough volume to be more than a testimonial page, a trade publication write-up you did not pay for, forum and community threads where customers describe outcomes in their own words, and comparison content that names competitors honestly enough to be credible.

None of that is fast. That is the point. There is no schema markup fix for a claim nobody else has ever repeated.

What to actually do

A sequence that fits inside a normal fortnight:

1. Pull the ten prompts that matter most to your business. Run each one in ChatGPT and Gemini separately, and record two columns: cited, and named.
2. Find the pages that are cited but never named. These are your ghost pages, and they are usually your best informational content.
3. For the top three, add a comparison or evaluation section that names real alternatives, including where you are not the right choice.
4. Pick one claim you make constantly and find or build one piece of third-party evidence for it that is not a testimonial on your own site.
5. Re-run the ten prompts in six weeks. Compare the named column, not the cited column.

Query phrasing matters too, and it cuts against most keyword habits. Short conversational prompts produced brand mention rates near 100% in the study, while long structured prompts on the same topic produced 2% to 3%. Test the short version of your question, because that is how people actually ask.

What can go wrong here

The obvious risk is over-correcting. If you decide informational content is worthless and stop making it, you lose the citations that build the model’s familiarity with your domain in the first place. Citations and mentions come from different places: citations reflect domain authority and original work, mentions reflect brand familiarity and positioning. You need both, and killing one to chase the other is how sites lose ground in both.

The second risk is measuring this too often. Answers vary run to run. A single check on a single day tells you almost nothing, which is exactly the trap the sample-of-one approach falls into. Six-week intervals, ten fixed prompts, two engines minimum.

And this may not be worth your time at all. If you sell into a market where buyers do not use assistants for discovery, if your pipeline is referral and relationship driven, or if you are a publisher whose business model runs on the citation link rather than the name, then citation rate is the correct metric and the ghost citation is not your problem. Check who your buyers actually are before rebuilding a content plan around this.

The one thing worth taking away: stop treating a citation as proof of visibility. It is proof your content was useful to a machine. Whether it was useful to your business is a separate question, and it has a separate column in the spreadsheet.

The full dataset is in the ghost citations study, and the trust argument is set out in this piece on AI search and branded content. If you want the other half of the picture, what these systems can and cannot read on your site, we covered that in how AI builds a B2B shortlist before anyone calls you.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post AI Search Is Citing Your Blog and Not Saying Your Name. Here Is What Changes That. appeared first on Mark8ng.com.

Google announced the change on 10 August. New AI Overviews sit at the top of the Google Analytics homepage summarising what shifted since you last logged in, with opt-in notifications by phone or email. The Google Ads homepage has been rebuilt around personalised insight cards with a prompt box above them. Dashboards in Google Ads turn a text prompt into a chart, and every report generates a real-time summary explaining the reason behind the numbers. Google Analytics gets benchmarking through Ask Advisor, comparing your campaigns against anonymised averages from similar businesses. All of it runs on Gemini, and all of it is in beta for English-language accounts.

Most of the coverage is treating this as a productivity story. It is, partly. But the two features people will actually rely on are the two that behave worst on small accounts.

The “why” is the risky part, not the automation

Automating a report is safe. If the chart is wrong, you can see it is wrong. Automating the explanation is a different thing, because an explanation arrives already interpreted, and the reader has no way to audit the reasoning that produced it.

Key driver analysis needs enough events to separate signal from noise. On an account running 20 to 40 conversions a month, almost any week-on-week movement is inside the range you would expect from randomness alone. Ask a system to explain that movement anyway and it will find something, because that is the job you gave it. Device mix shifted. A campaign’s impression share moved. One geo underperformed. All true statements. None of them caused anything.

Here is the version that will show up in real accounts. A plumbing firm with two Search campaigns sees conversions drop from 31 to 22. The summary says mobile conversion rate declined and suggests reviewing mobile landing page experience. The account manager spends two days on page speed. The actual cause was that a competitor started bidding in the same postcode on the Tuesday, or that nine people who would have called in week three called in week four instead. Nothing was fixed, because nothing was broken in the way the summary described.

The dangerous feature is not that the agent is confident. It is that the confident version is now the default view, and the raw report is one click further away than it used to be.

Where the benchmark quietly falls apart

Benchmarking against “anonymised averages from similar businesses” sounds like the number every small advertiser has wanted for a decade. The problem is that you cannot see the peer set.

You do not know how “similar” was defined, how many accounts are in the comparison, or whether the businesses in it are pursuing the same goal you are. A brand optimising for cheap leads and a brand optimising for qualified leads will sit in the same category and produce wildly different cost per conversion. If your number looks bad against the average, that may mean you are inefficient, or it may mean you are the only one in the set who bothered to filter out rubbish enquiries.

Treat the benchmark as a prompt to investigate, never as a target. The moment somebody in a meeting says “Google says we should be at £18 and we are at £27”, the number has done more damage than good.

What is safe to accept, and what is not

A rough division that holds up in practice:

• Reasonably safe to accept: disapproval diagnostics, budget pacing flags, obviously wasted spend, missing extensions, broken tracking. These are checkable facts, and you can verify them in under a minute.
• Verify before acting: keyword and asset suggestions, seasonal recommendations, structural changes. Reasonable ideas, generated without knowing your margins.
• Do not accept on the summary alone: any statement about causation, any benchmark comparison, any recommendation that changes audience or creative strategy.

The habit worth building now, while the features are still new enough that nobody is dependent on them: before you act on a “why”, open the underlying report and check whether the movement is even outside your normal weekly range. If you do not know what your normal weekly range is, that is the first thing to work out, and it matters more than anything the agent will tell you this quarter.

When this is genuinely useful

None of this means the tools are bad. On an account with real volume, thousands of conversions a month across multiple campaigns, driver analysis is doing something a human analyst would take hours to do, and doing it every morning. Agencies managing forty accounts will get a triage layer that tells them which four to look at today. That is worth having.

The gap is between the accounts where this works and the accounts where it produces confident fiction, and Google’s interface makes no distinction between them. If you already send your marketing data somewhere you can query properly, our note on Google’s new BigQuery marketing connectors covers the other half of this problem.

What to do this week

Log in and look at the cards, but do not act on them yet. Spend an hour instead working out the natural weekly variation in your conversion count over the past six months. Write the range down. That single number is what turns every future AI summary from an instruction into a claim you can test.

And if your account does not have the volume for any of this to mean much, the honest answer is that the most valuable feature in the release is the disapproval troubleshooting, and everything else is a reason to be more careful, not less.

Google’s own description of the rollout is in its Google Ads and Analytics AI update, and it is worth reading for the wording alone: the stated aim is to keep marketers “firmly in the driver’s seat”, which is a promise the interface design does not entirely support.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Google’s New AI Agent Reads Your Ad Account and Tells You What to Do. It Does Not Know How Little Data You Have. appeared first on Mark8ng.com.

On 8 August Google added connectors for Shopify, Klaviyo, HubSpot and Mailchimp to BigQuery Data Transfer Service, all in preview, alongside Microsoft SQL Server. PostgreSQL, MySQL and Snowflake moved to general availability. The Shopify connector pulls order histories, inventory logs and customer profiles. Klaviyo pulls email and SMS engagement logs including sends, clicks and opens. HubSpot syncs pipeline and contact tracking. Mailchimp brings campaign performance and audience list attributes.

Stated plainly: the Monday morning CSV job is now a configuration screen.

The part that actually changes the maths

Marketing teams have been able to warehouse their data for years. What stopped most small teams was not capability, it was the standing cost of a third-party pipeline tool sitting between the platforms and the warehouse, priced per row or per connector, on top of a warehouse bill.

Google’s pricing removes one side of that. Ingestion from first-party Google sources is free, which covers Google Ads, Google Analytics 4, Campaign Manager, YouTube and Cloud Storage, with Google Play the exception. Amazon S3, Azure Blob Storage, Redshift and Teradata are also free to ingest. The new third-party connectors run on consumption pricing rather than row counts, with compute in major regions quoted at under six cents per slot-hour. Google also quotes a monthly uptime target at or above 99.99%, though that figure describes the service rather than the preview connectors specifically.

For a twelve-person direct-to-consumer brand, that is the difference between a data project needing a budget line and a data project needing an afternoon. Ad spend and site behaviour arrive free. Orders and email engagement arrive cheaply. Everything lands in one place with the same customer identifiers available to join on.

Getting the data in is the easy half

This is where these announcements usually oversell themselves, so it is worth being blunt. Loading four sources into a warehouse does not produce an answer. It produces four tables that still disagree with each other, now in SQL instead of Excel.

The work that remains is identity. Klaviyo knows a person by email. Shopify knows them by customer ID, and by a different one if they checked out as a guest. HubSpot has a contact record that may or may not match either. Somebody has to decide which identifier is authoritative and write the logic that stitches them, and that decision is a business judgement, not a technical one. Nothing in this release makes it for you.

Then someone has to write the queries. If nobody on the team can write SQL, a warehouse full of clean data is an expensive filing cabinet. The honest version of the plan includes either a person who can query it or a reporting layer on top, and both cost something.

What can go wrong

Preview status is the first thing to respect. Preview connectors can change schema, behave inconsistently, or be withdrawn. Do not build the report your board reads on a preview connector, and do not switch off the existing process until you have run both side by side for a full month and reconciled the difference.

Cost surprises are the second. Consumption pricing is cheap when volumes are modest and stops being cheap when someone schedules a full historical resync every hour because incremental transfers looked complicated. Set the schedule deliberately, check the first month’s bill against expectations, and set a budget alert before you set anything else.

The third is the one that gets people in real trouble. A Shopify customer profile and a Klaviyo engagement log contain personal data. Moving them into a warehouse creates a new copy in a new place, and your retention policy, deletion process and data map all need to know it exists. If a customer asks to be deleted, the request has to reach BigQuery too. Deciding that afterwards is much harder than deciding it now.

And a quieter one: a warehouse does not solve attribution. It gives you the raw material to build a view. The interpretation problems do not go away, in the same way that a new metric appearing in Search Console does not tell you what it means.

When to leave this alone

If your business runs almost entirely inside one platform, the native reporting is usually enough and this is a solution looking for a problem. A Shopify store doing most of its revenue through Shopify’s own checkout, with email as a supporting channel, can answer nearly every useful question without a warehouse. The case gets strong when you have three or more systems that each hold part of the customer, when the same question gets a different answer depending on who you ask, or when someone is spending a recurring half-day per week on exports.

The signal worth watching in this release is not the connector list. It is that ingestion is being priced towards zero while the value moves up the stack, into modelling, into the agent layer, and into whoever can ask the warehouse a sensible question. Getting your data in is becoming a commodity. Knowing what to ask it is not, and that is where a small team should be spending its next hour.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Google Just Made Marketing Data Warehousing a Checkbox. That Is the Easy Half. appeared first on Mark8ng.com.

That is what shadow AI actually looks like in a marketing team. Not espionage. A deadline.

The scale is not small. A BlackFog study on unsanctioned AI use found 49% of respondents using AI tools their employer had not approved, and 60% saying they would take risks to meet a deadline. Other surveys land anywhere between 50% and 80% depending on how the question is asked. The exact figure matters less than the consistency of the direction, which is that roughly half your team is doing this and you probably have no list of what they used.

The diagnosis worth borrowing, and the prescription worth questioning

A piece published this week on CDOTrends makes a useful argument: shadow AI in channel marketing is a signal of unmet workflow need rather than a discipline problem. Channel work spans partners, distributors, regional teams and agencies, which produces constant demand for drafts, localisation, summaries and approvals. When the approved path is slow, people take the fast one. The author’s phrasing is that employees see two choices, wait or solve it themselves, and the job is to offer a third.

That framing is right, and it is worth saying plainly that the article making it was written by a growth marketing director at a company selling channel marketing automation with governance built in. That does not make the diagnosis wrong. It does mean the conclusion arrives conveniently at “buy a governed platform,” which is not an available answer for a six-person agency or a founder-led marketing team. The transferable part is the risk tiering. The platform part is not.

Tier by data, not by tool

Most small teams that attempt an AI policy start by listing approved tools. That list is out of date within about six weeks, and it answers the wrong question anyway. What creates exposure is not which product someone opened. It is what they put into it and whether anyone read the output before it reached a customer.

Three tiers cover almost everything a marketing team does.

Low risk. Anything built from information that is already public. Reworking approved campaign copy, drafting from a published case study, summarising a webinar you streamed publicly, generating subject line variations. No approval needed, no logging needed. Say so explicitly, because ambiguity here is what pushes people into hiding the higher-risk work too.

Moderate risk. Partner and client communications, performance summaries, translation and localisation of assets, anything referencing a named account. Approved tools only, output read by a second person before it leaves the building.

High risk. Pricing, lead scoring, customer segmentation, contract language, anything containing a customer list, and anything under an NDA. Named owner, written record of what went in, and a default answer of no unless the tool is covered by an agreement you could show a client.

A concrete version for a small agency: client names and unreleased campaign material are high risk, agency blog drafts are low risk, and the monthly client report sits in the middle because it contains their numbers but not their customers. Three sentences. That is a policy people will actually follow.

The inventory question that gets a real answer

Asking a team to declare which AI tools they use produces a tidy list and a false sense of coverage, because the answer is filtered through whether people think they are in trouble. Ask a different question instead: in the last month, what is the most sensitive thing you have pasted into an AI tool, and did anyone check the output before it went out?

You will get better information if the stated consequence of answering honestly is guidance rather than a telling-off. That is the single most useful line in the CDOTrends piece and it costs nothing to adopt.

What goes wrong

Blanket bans are the classic failure. Prohibit the tools and the usage does not stop, it moves to personal devices and personal accounts, where you have no visibility at all and no way to retrieve anything. You have traded a manageable problem for an invisible one.

Over-governance is the quieter failure. A four-person team does not need an intake form, a risk-rating committee and a quarterly review cycle. Build that and nobody will use it, which returns you to shadow AI with extra paperwork. Scale the process to the size of the team, not to the size of the risk you read about in an enterprise article. The same judgement applies when deciding what to hand over to marketing agents, where the question is scope rather than paperwork.

The third one is assuming a paid business tier solves it. A business subscription changes the data handling terms. It does not stop someone pasting a client’s customer list into a chat window, and it does not create the record of what went in. The control you actually need is a habit, not a plan.

When this is not worth doing

If you are a solo operator working on your own material, with no client data, no NDAs and nobody else touching the tools, writing a policy for yourself is theatre. Spend the hour somewhere else. This becomes real the moment a second person handles someone else’s information on your behalf, which for most businesses is the first contractor.

The genuinely uncommon takeaway here is that a shadow AI problem is a diagnostic. Every unapproved tool in your team is a place where the sanctioned way of working was too slow to use. Before writing rules, look at what people reached for and ask why the approved route lost. Fix that, and half the policy stops being necessary.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post Your Team Is Already Using AI You Did Not Approve. The Tool List Is Not the Problem. appeared first on Mark8ng.com.

On 6 August OpenAI said it is removing limits on text chats and making GPT-5.6 Luna the default model for Free and Go accounts, replacing GPT-5.5 Instant. Free and Go users also get a Think button that routes harder questions through more reasoning. Plus and Pro accounts move to an updated GPT-5.6 Sol with a slider controlling how much thinking goes into an answer. Limits on files, images, voice and image generation stay in place. The announcement arrived shortly after ChatGPT passed a billion weekly users, with unlimited text chats for free accounts scheduled to land the following week.

The number worth holding onto is not the billion. It is the split.

Two tiers, two models, two answers about you

Until now, free accounts had a ceiling. Hit it and the conversation stopped, or dropped to a weaker fallback. That ceiling quietly rationed how many people used ChatGPT for the kind of open-ended question that mentions brands: which accounting tool suits a sole trader, who does commercial kitchen servicing near Leeds, what is the difference between two CRMs. Remove the ceiling and that behaviour has nowhere to stop.

So the model answering most brand questions is now GPT-5.6 Luna, and almost nobody in marketing has tested against it. OpenAI’s own evaluation puts factual errors 62% less common in Luna than in GPT-5.5 Instant, and 68% less common in Sol. Take those as directional rather than settled, since they are internal numbers with no published methodology. But the gap between the two figures is the point. Luna is better than what free users had. It is still not the model you are testing in.

A worked example. A regional B2B supplier checks how ChatGPT describes them, using the founder’s Plus account with the thinking slider left high. The answer is detailed, cites two trade publications, mentions a certification they hold. Reassuring. A buyer on a free account asks the same thing, gets a shorter reply from a different model, and the certification does not appear because the source it lives on was not retrieved. Nothing about the company changed between those two answers. Only the tier did.

What to actually do this week

Run the check twice, from two accounts, and write down which model produced each answer. That is the whole method. It takes about twenty minutes and it is the only way to see the gap.

Three questions are usually enough: your brand name alone, your main category with a location or qualifier attached, and a comparison question naming a competitor. Ask each in a fresh conversation. Log the date, the tier, the model, and whether any link was offered. Repeat monthly rather than daily, because the answers move around enough that a single day tells you very little.

If the free-tier answer is thinner, look at what the paid answer used that the free one did not. Usually it is a source that takes more retrieval effort to reach: a PDF, a page buried three clicks deep, a claim that only exists on a third-party directory. The fix is rarely clever. It is putting the fact somewhere shallower.

What can go wrong here

Three failure modes, and the first is the common one.

Treating a handful of answers as data. You are running a sample of two on one afternoon. Chatbot output varies between sessions for reasons that have nothing to do with your website. If you rewrite a page because of one bad answer, you will be rewriting it again next month. We covered the broader version of this problem in how AI describes your brand across many queries rather than one, and the same caution applies here.

Chasing model changes as a strategy. Default models will change again. Building a workflow around Luna specifically is building on something with a short shelf life. What lasts is the habit of checking both tiers, not the model name.

Assuming unlimited means used. Removing a cap raises the ceiling on usage. It does not prove your buyers were hitting that cap in the first place. If you sell into a market that researches through trade bodies, procurement portals or a rep they have known for years, none of this may show up in your pipeline for a long time.

When this is not worth your time

If your business runs on repeat customers and referrals, and search has never driven meaningful revenue, this is a twenty minute curiosity rather than a project. The teams that should take it seriously are the ones where a stranger comparing three options is a normal way to get discovered. For everyone else, knowing the gap exists is enough for now.

The uncomfortable part of this update is not the model change. It is that the AI answer most of your market sees has always been a different answer from the one you have been reading, and the tier that produces it just got a much bigger audience. Check both. Then decide whether the difference is worth fixing.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

The post ChatGPT’s Free Tier Just Went Unlimited. You Have Been Checking the Wrong Model. appeared first on Mark8ng.com.

On 6 August a New Mexico judge, Bryan Biedscheid, ordered Meta to pay $567 million into an abatement fund, taking the company’s total liability in the case to $942 million. He also ordered a set of product changes, and those are the part worth reading if you buy social advertising. Users under 18 in New Mexico get their monthly usage capped at 90 hours, push notifications paused between 10pm and 7am, and Like counts hidden unless a parent or guardian approves showing them. Meta was further ordered to keep improving age estimation from behavioral signals and to build a dedicated under 13 prediction model within two years.

A curfew is more interesting than a fine

Advertisers have watched large platform fines land for a decade and correctly concluded they are a line item for the platform and nothing else. This one is different because the remedy is a product change. The court found the conduct amounted to a public nuisance and ordered the harm reduced rather than the service stopped, which means the fix is engineering rather than accounting. The state’s own summary of the ruling reads like a document written to be copied.

A notification blackout from 10pm to 7am removes one of the more reliable evening engagement spikes for the under 18 audience in that state. A 90 hour monthly cap removes inventory outright. Neither will show up in a national campaign report, because New Mexico is about 0.6% of the US population. That is exactly why it is easy to miss, and exactly why it matters.

The template problem

New Mexico’s attorney general has said he intends to introduce a state social media safety bill carrying an age verification mandate, and legislators including Rep. Linda Serrato are preparing a wider package for the 60 day session that starts in January. Other states are watching the same result.

If that plays out, the operational problem is not any single rule. It is fifteen slightly different rules, with different age thresholds, different quiet hours and different disclosure requirements, arriving over two years. Building a process for one state now is cheaper than retrofitting one for fifteen later.

What a brand can do this quarter

Take a skincare brand whose paid social skews 16 to 24. The useful work here is not legal, it is knowing your own numbers. What share of your reach is under 18, broken out by state? What share of your conversions land between 10pm and 7am? Without those two figures you will not be able to tell a policy effect from a weak creative when the next state acts.

The second piece is a decision you only make once: whether you want under 18 reach at all. Plenty of brands carry it passively because the platform delivered it, not because anyone targeted it. Excluding it costs some cheap impressions and removes a category of risk that is clearly growing.

Where this goes wrong

For most advertisers the overcorrection is worse than the exposure. Age signals on social platforms are estimates, not documents, and tightening exclusions aggressively will also cut the 18 to 24 year olds who are your actual customers. A court order aimed at platform design is not an instruction to shrink your own audience by a decade.

Assuming the platform’s compliance covers yours is the other error. The order binds Meta. It says nothing about your creative, your data collection, or the disclosure rules that already apply to synthetic performers in your ads.

Editor’s note: This area changes quickly, so check the latest platform policy before making compliance decisions.

Pull an under 18 reach breakdown by state from your last quarter of Meta reporting this week. It takes ten minutes, and it is the baseline every future version of this question gets measured against.

The post A Court Just Redesigned Instagram for Teenagers. Your Ads Run Inside That Redesign. appeared first on Mark8ng.com.