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.

One survey put the share of UK senior decision makers using ChatGPT, Copilot or Perplexity somewhere in their procurement process at 66%. Treat the exact number with caution, because self reported tool usage surveys run high almost by design. The direction is harder to argue with. The stretch of the buying process that used to take a week of searching and comparison now happens in a single conversation, and it ends with a shortlist and a recommendation on who to call first.

The failure mode is not bad content

Suppliers that surface reliably are the ones whose credentials sit in text a machine can read. Suppliers that do not surface are often just as capable, with the same certifications and better outcomes, filed in PDFs, gated portals and slide decks. Analysts describing AI as the first gatekeeper in procurement keep landing on the same point: the gap is structural, not qualitative.

Take a mid sized industrial supplier with fourteen years of certified process and a customer list any competitor would envy, all of it inside a capability brochure behind an email form. To a model building a shortlist, that company has a homepage, a contact page and nothing to say. The competitor with a weaker product and a plain HTML certifications page takes the slot.

Four things to move into text this week

Most exclusions trace back to the same four gaps. Pricing, or at minimum a pricing model and a range. Integrations, named individually rather than shown as a logo wall image. Certifications and compliance, with the issuing body and the year. Case study outcomes, with the number written in the sentence instead of sitting in a downloadable file.

None of that needs new content. It needs existing content moved out of assets and onto pages. The test takes five minutes: open your site with images and JavaScript disabled and read what is left. That is roughly what the model gets.

The cost that comes with being findable

Being easier for AI to find also means being easier for AI generated noise to find. Suppliers are already reporting inbound inquiries routed automatically to every name on a category list, arriving in a form that looks like qualified interest and is not. A two person sales team can lose a week to that before anyone spots the pattern.

Decide the filter in advance rather than after. No company domain, no specific requirement, and a timestamp that matches several competitors to the minute is usually the whole tell.

When this is not your problem

If your pipeline runs on referrals, or you sell into a market where four buyers know every supplier by name, none of this changes your quarter. Machine readability matters most when a buyer does not already know who you are, which is another way of saying it is a top of funnel problem. If you do not have one, spend the time elsewhere.

The risk on the other side is treating this as a content project. Publishing twelve more blog posts does not put you on a shortlist if your pricing is still a contact form. The order matters more than the volume.

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

Write down the five questions a buyer would ask an assistant before shortlisting a company like yours, then run them. If you are not named, compare what the named companies publish in plain text against what you publish in a PDF. If you want a data source rather than a manual log, Microsoft now reports which topics AI links to your brand. That gap is usually the entire answer.

The post AI Builds the Shortlist Before Anyone Calls You. Here Is What It Can Read. appeared first on Mark8ng.com.

The figure surfaced in reporting on Time, which started selling ads inside its own markdown pages last month. We already covered how media buyers reacted to that ad product. The more useful question for a company that will never buy a Time placement is the one underneath it: does your site need a machine readable twin at all?

What a markdown twin actually is

Markdown is text with light formatting. A heading is a hash mark, a link is brackets and parentheses, everything else is stripped out. A language model parsing a markdown file spends no effort on your cookie banner, your mega menu, your lazy loaded hero image, or the nine scripts that fire before your first paragraph. It gets the words.

The theory is that easier access means more accurate retrieval, and more accurate retrieval means a model is likelier to describe you correctly and cite you. Time went a step further and sold ad space inside those files, formatted as FAQs and labeled as sponsored. Ally Bank and the Project Management Institute bought the first units. Digiday reported that Time now sees more bot traffic than human traffic on most days, which is the actual business reason the product exists.

The part nobody can prove yet

There is no published evidence that a markdown twin increases how often a brand gets cited in AI answers. Not contested evidence. No evidence, because nobody has run the study. The publishers building these are hedging against falling search referrals, which is a different motivation from a brand trying to get named in a buying conversation.

There is also a risk worth naming out loud. If your markdown version says something different from your HTML version, you have built cloaking. Serving one thing to crawlers and another to people is the exact pattern search engines have penalized for twenty years, and no AI company has published a rule on it yet. Rob Derow at BCG X raised the same concern about Time. A gap that starts as a content management oversight looks identical, from the outside, to deliberate manipulation.

The version that costs nothing

Take a seven person B2B software company with a pricing page, four case studies and an integrations list. Before building anything new, check what a crawler gets from what already exists. Are the case study outcomes in HTML text, or in a gated PDF? Is the pricing a number, or a button that says Contact Sales? Is the integrations list a page, or a JavaScript widget that renders after load?

On most small sites the machine readability problem is not the file format. It is that the useful facts are locked inside assets a crawler cannot open. Fixing that takes a few afternoons and improves the human version of the site at the same time. Building a parallel markdown tree costs maintenance forever, and maintenance is exactly where the cloaking risk lives.

When it is worth building

A markdown twin starts to make sense if you publish at volume, if your server logs already show heavy AI crawler traffic, and if the two versions can stay in sync automatically rather than by hand. Publishers meet all three conditions. A company shipping four blog posts a month meets none of them.

This may not be worth your time if you cannot answer one question: how many AI crawler requests hit your site last month? That number sits in your server logs or your CDN dashboard. If it is small, you are optimizing for an audience that is not visiting.

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

Pull last month’s crawler hits by user agent before you scope anything. If GPTBot, ClaudeBot and PerplexityBot barely appear, your problem is not file format, and a second website will not fix it.

The post Brands Are Building a Second Website for Machines. Yours Probably Does Not Need One. appeared first on Mark8ng.com.

How bad the slop problem actually got

Data from AI detection firm Pangram, reported by TechCrunch, found that more than 40% of long-form LinkedIn posts are now fully AI-generated, and that LinkedIn accounts for roughly 62% of all AI content flagged across the major social platforms Pangram scanned. That is the backdrop for the new “Seems like AI slop” button, reachable from the three-dot menu on any post. Flagging a post hides it from your own feed immediately and feeds a signal into LinkedIn’s internal classifiers. LinkedIn is also quietly scaling back its own writing assistant, replacing the old “Enhance your post” rewrite tool with a more conservative proofreading feature built to preserve a person’s actual voice instead of smoothing it into the same voice as everyone else’s.

What Collaborative Posts actually changes

Collaborative Posts is the other half of the strategy. According to LinkedIn’s own documentation, only the account that created the post can edit it or manage who is listed as a co-author, and every invited collaborator has to actively accept before their name appears. Posts have to be public to qualify, and Company Pages added as collaborators do not receive any engagement metrics at all, only the individual profiles do. That last detail matters more than it sounds: a brand can co-author a post with an employee or executive, reach that person’s network through it, and still not be able to see how it performed. The Page’s own analytics dashboard is the only place that data shows up.

The out-of-network metric nobody is using yet

LinkedIn added an in-network versus out-of-network breakdown to post analytics back in June, splitting impressions between your existing audience and everyone reached beyond it. AJ Wilcox, writing for Social Media Examiner, recommends tracking which topics and formats consistently travel past your own follower base rather than staying contained within it. That is a more useful optimization target than raw engagement, because it tells you which content is actually earning new reach instead of just getting liked by people who already follow you.

A workflow worth trying, and where it breaks

A small agency running a founder’s personal LinkedIn presence could reasonably start co-authoring one post a month with a client executive using Collaborative Posts, then check the out-of-network number a week later to see whether the joint post reached further than either account does alone. Do this for a quarter and you will have real data on whether co-authorship earns reach or just adds names to a post that would have performed the same either way. Where this breaks down is scale. The five-collaborator cap and the requirement that every collaborator actively accept make this a slow, deliberate format, not something you can run across twenty client accounts at once. If your LinkedIn strategy depends on volume, this feature will not help you, and trying to force it will look exactly like the kind of manufactured authenticity the AI slop button was built to catch.

Editor’s note: This area changes quickly, so check LinkedIn’s current help documentation before building a workflow around Collaborative Posts, since parts of the rollout were still marked early access in LinkedIn’s own support pages after the global announcement.

If you are also cleaning up AI-generated content across other platforms, we covered the wider crackdown when Snapchat, Substack, and LinkedIn all started flagging AI content in the same week.

The post LinkedIn Capped Collaborative Posts at Five and Added an AI Slop Button. Here Is the New Playbook. appeared first on Mark8ng.com.