Search Console AI Report Opens, ChatGPT Ads Misfire

Estimated reading time: 28 minutes

Measurement finally arrived this week, and then stopped just short of being useful. Google’s generative AI performance report is now live in every Search Console property, though it still withholds clicks and queries. Elsewhere in this week’s search marketing news: SE Ranking put ChatGPT ads under a 50,000-prompt microscope and found one in seven landing on the wrong topic entirely, Meta appears to be quietly building its own web index, Google Ads and Analytics grew a conversational AI layer, and a joke file about cats exposed how flimsy the evidence behind a popular GEO tactic really is.

Search Console's Generative AI Report Reaches Every Property

The generative AI performance report is now live across every Search Console property, according to Barry Schwartz, who checked all of his own profiles after a reader flagged it. Google hasn’t formally announced the full rollout, and a pop-up now points property owners straight to the report when they log in. It sits as an expandable tab beneath the main performance report.

Here’s what you get:

Impressions: how often your URLs appeared in generative AI features across Search and Discover
Pages: which specific URLs are surfacing inside AI features
Countries and devices: where and how people saw you
Dates: hourly, daily, weekly and monthly granularity

And here’s what you still don’t get: clicks or queries. Neither has appeared in the report since it launched in June.

That absence matters more than the rollout. When Google’s head of Search said last month that AI features send billions of clicks to websites every week, the obvious reply was that publishers cannot verify any of it at site level. This report could have settled the argument. Instead it hands you an impression count and asks you to take the rest on trust. Use it for what it genuinely offers, which is a directional read on which pages the AI surfaces favour, and pair it with your own analytics segmentation for anything resembling a traffic answer. The gap between the two numbers is the story your clients will keep asking about.

Google Tests a Homepage Without the Search Button

Google has confirmed a limited test that adds Create images, Ask about files and Brainstorm buttons beneath the search box on its homepage. Search Engine Watch, which spotted the test running for signed-out users, reports that the classic Google Search button disappears in the variant they saw, replaced by that trio, with the AI Mode button sitting inside the search field itself.

The homepage has barely changed in a quarter of a century. A logo, a box, two buttons. Replacing one of those buttons with prompts to generate images or interrogate your own files is not a cosmetic tweak, it’s a statement about what Google now thinks the box is for.

For search marketers the implication is upstream of any ranking factor. Every one of those three buttons routes a user into a surface where the answer is generated rather than listed, and where your website is a possible ingredient rather than the destination. It’s a test, and tests get abandoned. But the direction of travel has been consistent enough this year that planning for a homepage which nudges people away from the ten blue links is safer than assuming the search button survives.

The Engineer Who Built Google's AI Foundations Walks Out

Jeff Dean has left Google to found a new AI company called Discovery Loop. Sundar Pichai announced that DeepMind co-founder Demis Hassabis becomes Chief Scientist of Alphabet and Chair of Google DeepMind, while Koray Kavukcuoglu steps up to Senior Vice President of Google DeepMind with day-to-day leadership of the Gemini models.

Most SEOs have never heard of Jeff Dean, which is a bit like never having heard of the person who invented the road. He co-invented TensorFlow. He co-authored MapReduce in 2004, which seeded the entire big-data industry, and Bigtable in 2006, which shaped how web indexing scales across thousands of servers. His 2015 paper on knowledge distillation is the technique now used both to build efficient models and, less charmingly, to reverse-engineer rivals’ AI search results.

Succession dramas at Google are usually noise for practitioners. This one is worth filing because of who replaces him: putting DeepMind’s leadership in charge of Alphabet’s scientific direction confirms that research and Search are converging into one organisation rather than two. When the people building the models also set the direction for the product that ranks your pages, expect the pace of change in search surfaces to track model release cycles rather than traditional product roadmaps.

Hreflang Alternates Were Never Really Indexed

Gary Illyes has clarified something that trips up international SEO teams constantly: hreflang alternate URLs are not indexed “in the proper sense”. Google stores them as alternates against the canonical, which is why they show up as not indexed in Search Console while still doing their job perfectly well.

The clarification came from a LinkedIn thread where a practitioner asked why Google picks one language version as canonical and reports the others as not indexed. If you manage multilingual sites, this is the answer to a question that has generated countless pointless audits, and it belongs in your client FAQ documents rather than your issues backlog. It also chimes with the broader message from Google this summer about Search Console reports never having been a checklist. A URL sitting in an excluded bucket is often the system working exactly as designed.

Business Profiles Ban Repeated Bilingual Names

Google has added a rule to its Business Profile guidelines: you can no longer repeat the same business name in multiple scripts or languages. A restaurant that listed its name in both English and Mandarin, or a business duplicating itself in Latin and Cyrillic characters, now falls foul of the naming policy.

Businesses in genuinely bilingual communities did this for good customer-facing reasons, so this will feel harsh to some. It is also, transparently, a keyword-stuffing loophole being closed. If you manage local listings for multicultural areas or international clients, audit business names before Google does it for you; suspension over a naming violation is a far more expensive problem than editing a listing today.

WordPress 7.0.3 fixes twelve vulnerabilities, and one researcher described the nastiest as a login screen cross-site scripting flaw. High severity, and on the single page every administrator visits.

Update the sites you manage. That’s the whole action item.

Related and worth watching if your clients run large plugin stacks: a WordPress 7.1 accessibility improvement may break plugins that modify admin screens, so test staging environments before that release lands. Security patches and accessibility changes arriving in consecutive releases is a reminder that “we’ll do the updates next quarter” is a policy with a shelf life.

A Popular Forms Plugin Faces a Backdoor Accusation

An accusation circulated this week that WPForms Lite adds a backdoor to new installations. Roger Montti at Search Engine Journal did the sensible thing and installed it to find out, and reports the outcome was not what the accusation implied.

Two lessons sit inside this small story. The first is practical: if a plugin on your client sites gets publicly accused of something, check the claim before you rip it out, because emergency removals break forms and lose leads. The second is about how quickly plugin panic now spreads through the WordPress ecosystem, where a single viral post can generate more damage than most actual vulnerabilities. Read the testing, not the headline.

ChatGPT Serves Ads on a Quarter of Commercial Prompts

Sam Altman once called advertising combined with AI “uniquely unsettling” and a last resort. Less than two years later, SE Ranking has measured what that last resort looks like at scale: 50,006 commercial prompts across 20 niches, the same prompt set used for their AI Mode study last month.

ChatGPT returned an ad on 25.94% of commercial prompts, remarkably close to AI Mode’s 29.45%. The two platforms are monetising commercial intent at broadly the same rate, though the format differs: ChatGPT offers a single ad slot per answer, where AI Mode showed two competing ads 71% of the time.

The uncomfortable finding is relevance. About 14.35% of ads were no more related to their query than an ad paired with a random prompt from the dataset, and in News & Politics and Relationships more than half were off-topic. The researchers note their manual review found additional ads that passed the similarity test while still missing user intent, so the real mismatch rate is probably worse than one in seven.

Three other numbers deserve a place in your planning. YMYL exposure is inverted between the platforms: ads appear on 28.69% of healthcare prompts in ChatGPT against just 2.64% in AI Mode. Advertiser concentration is extreme, with BestMoney alone accounting for 13.56% of all ads observed, Hungryroot taking 79.92% of Food & Beverage and Booking.com 62.57% of Travel. And paying still doesn’t buy you into the answer: in 96.37% of placements the advertiser was not among the cited sources, an even starker version of what AI Mode showed. SE Ranking also ran their own campaigns, generating 97,000 impressions and 1,263 clicks over two weeks, with very few sign-ups. Early-mover advantage is real here, but so is the risk of paying for impressions next to questions your product cannot answer.

Meta Is Crawling the Web, and May Be Building a Search Engine

Three weeks ago we covered DataDome’s finding that Meta’s crawlers had made around nine billion requests in a quarter while returning essentially no referral traffic. This week, a plausible explanation surfaced. Pieter Levels, citing Meta staff who contacted him, posted that Meta is allegedly building its own search engine so that when its AI performs a web search it doesn’t route through Google and hand Google the training signal.

It’s an allegation rather than an announcement, and Meta has form here: the company wanted its own search service for over a decade, partnered with Bing, then dropped the partnership. But the logic holds better now than it ever did. If your AI needs a web index and every existing index belongs to a competitor, building your own stops being a vanity project and becomes infrastructure.

So the nine billion requests were not waste after all; they were construction. That reframes the crawler-blocking decision we discussed last month. Blocking Meta’s bots today denies training data to a company that gives you nothing back, but it may also exclude you from an index that could carry meaningful discovery traffic within two years. There’s no obviously right answer, only a decision that should be made deliberately and revisited quarterly rather than left to whatever your CDN defaults to.

A File About Cats Passed Every Test Used to Sell LLMs.txt

This is my favourite piece of the year so far. Mark Williams-Cook grew tired of watching the industry treat “an AI bot fetched it” and “ChatGPT said it helps” as evidence that llms.txt does anything, so he invented a competing standard called cats.txt: a plain-text file declaring your office cats, their job titles, breeds, and a mandatory affection metric called PurrLevel, scored out of ten. He wrote a straight-faced specification, published it, and posted about it on LinkedIn, because as he notes, large language models hold LinkedIn in unaccountably high regard.

Then he ran the file against the exact four proofs the industry cites for llms.txt. AI bots crawled it. Google indexed it. LLMs returned details about a cat that exists nowhere except in that file. ChatGPT confirmed at length that cats.txt could help him rank.

“None of which is evidence of anything, which was rather the point.”
– Mark Williams-Cook, Director at Candour

The joke escalated beautifully, with other SEOs adding cats.txt files to their own sites and someone building a cleaner, better-specified version at catstxt.org. Williams-Cook is careful to say he isn’t claiming llms.txt will never work; his point is that the evidence bar currently used to sell it was cleared without effort by a file about a tuxedo cat called Odd.

Put this alongside John Mueller’s comments last month about markdown mirrors being technical debt, and a pattern emerges that should shape how you buy and sell GEO services. Crawled is not the same as used. Indexed is not the same as influential. An LLM repeating your content proves retrieval, not ranking benefit. Before the next tactic goes on a client invoice, ask what result would look different if the tactic did nothing at all. If you cannot answer, you are selling astrology.

50,000 Query Fan-Outs, and a Free Dataset to Go With Them

Moz has published a dataset of 1,000 subtopics and 50,000 fan-out prompts, built to model how AI systems expand a single query into many, and released the raw data free.

The finding practitioners should sit with concerns brand bias. When Moz’s early prompt generation mentioned brands or competitors directly, the odds of those brands appearing in the LLM response were close to 100%, which is a spectacular way to fool yourself in a visibility audit. For the published dataset they used brands as contextual hints instead, a nudge rather than a command, and brand mentions dropped to 12.8% of generated prompts. More usefully still, just two fan-out types accounted for 97% of all brand mentions: Entity and Comparison prompts.

That last number is directly actionable. If comparison-shaped questions carry most of the brand visibility in fan-out, then comparison content, the honest kind that names alternatives and explains trade-offs, is doing double duty in AI search. It also means any AI visibility tracking that leans on brand-loaded prompts is measuring its own prompt design. Check how your tools generate their prompt sets before you present the outputs as a competitive benchmark.

Clarity Splits Branded From Non-Branded AI Citations

Microsoft Clarity’s AI Citations dashboard now labels grounding queries as branded or non-branded, adds filters for each, and reports Share of Authority separately for both groups. Clarity remains free, though an administrator must verify the domain first via the tracking code, Search Console or Bing Webmaster Tools.

The distinction is more useful than it first appears. A grounding query is what the AI system actually searched for behind the scenes, which frequently differs from what the user typed. Branded retrieval means the system went looking for you by name. Non-branded means your content surfaced while the AI researched a category or problem, which is the only one of the two that represents genuine discovery.

One caution on the headline metric: Microsoft calculates Share of Authority only across the queries where your domain already earned a citation, so it shows how much of that conversation you owned, not your share of AI search overall. Reported carelessly to a client, it will look far more flattering than it is. Track citations, AI referral traffic and conversions as three separate lines, because more citations do not automatically become visits, and visits do not automatically become revenue.

AI Assistants Become Consumer Brands While Trust Sits at 28%

YouGov’s latest BrandIndex analysis has Claude’s consideration score among US Gen Z adults nearly doubling in Q2, from 14.2% to 28.1%, the highest score of any brand it ranked. OpenAI more than doubled too, from 10.1% to 22.1%, and ChatGPT tracked separately grew from 16.3% to 25.4%. Meanwhile only 28% of Americans say they trust AI assistants for answers.

Greg Jarboe’s write-up is admirably sceptical about its own headline, and the caveats are the valuable part. Google Assistant’s score also climbed, from 8.1% to 15.8%, during a quarter when Google was actively steering users away from Assistant towards Gemini, which looks like residual name recognition catching a lift rather than any strategy working. And the same ranking featured Johnnie Walker, REI and Dove Baby posting comparable gains on the back of ordinary campaigns and seasonal promotions.

Strip the generational framing away and one fact survives: AI companies now appear in consumer brand-tracking datasets alongside whisky and outdoor retailers. For GEO planning, the trust gap is the number to hold onto. People increasingly consider these assistants, yet most still verify what they’re told, which means your brand’s job is not only to be named in the answer but to survive the check that follows.

Yelp Puts Restaurant Bookings Inside ChatGPT

ChatGPT users can now reserve a table or join a Yelp waitlist without leaving the conversation. Discovery to booking, in one thread.

The pattern is now unmistakable across both major assistants. Google spent last month wiring Instacart, Canva and YouTube Music into AI Mode; OpenAI is doing the same with the transactional layer of local search. In both cases the winner is the platform that holds the inventory and the booking rail, and the loser is whoever assumed the customer would arrive on their website to convert.

For hospitality clients specifically, that makes your presence in aggregator inventory a search visibility question rather than a distribution afterthought. If your restaurant’s availability is not in the system feeding the assistant, you’re not in the answer, however good your own booking page is. Worth reviewing which aggregators actually carry your clients’ live availability, because those relationships now determine whether an AI conversation can end in a booking at your venue or somebody else’s.

Hidden Text Can Poison an AI Agent's Memory for Months

Forcepoint’s X-Labs has published a threat model for what it calls persistent memory poisoning, and the example should make anyone running an agentic workflow uneasy. An assistant with browser access reads a page about travel disruption. Hidden near the bottom, sized so no human will ever see it, sits a paragraph asserting that a particular company is the official emergency booking provider. The extractor doesn’t distinguish hidden text from visible text, so the claim gets filed as a fact. A month later the user’s flight is cancelled, they ask their assistant what to do, and out comes the recommendation.

I’ll be honest, this one unsettled me more than most security stories, because the attack needs no privileged access at all. The underlying academic work, MINJA from NeurIPS 2025, poisons agent memory purely through ordinary queries submitted via the standard interface, and it has been demonstrated against ChatGPT, Gemini, Claude and Microsoft 365 Copilot.

Two implications for our industry. Defensively, if your clients deploy agents that browse the open web, the pages those agents read are now an attack surface, and hidden text on third-party sites can shape recommendations weeks later. Offensively, and let’s be blunt about where this leads: the tactic is a short hop from the hidden-text spam that search engines spent twenty years learning to detect. Anyone selling “get recommended by AI” services with suspicious confidence is worth asking exactly what technique sits behind the promise.

Google Ads and Analytics Grow a Conversational AI Layer

Google is expanding Ask Advisor with agentic capabilities across both Google Ads and Google Analytics. Analytics gains AI Overviews on the homepage summarising what changed since you last logged in, with optional email and mobile notifications. Google Ads gets a redesigned homepage carrying personalised AI insight cards, where you can ask natural-language questions such as how competitors are affecting your impression share. Both platforms are getting Dashboards built from text prompts.

Separately, Google announced that Ask Advisor will benchmark campaign performance in Analytics against anonymised averages from similar businesses, which is the feature most likely to end up in a client conversation. Benchmarking against your own history is analysis; benchmarking against an unnamed peer set is a number you cannot interrogate.

Two practical notes. First, an AI summary that tells you what changed is genuinely useful for account managers juggling many accounts, and it will surface things a weekly check might miss. Second, every insight is generated by the platform that profits from your spending decisions, which is the same caveat that applied to the missed-growth estimates in the Recommendations tab last month. Treat the summaries as a prompt for investigation rather than a conclusion, and keep the analysis that informs budget changes in a tool that isn’t also the vendor.

AI Max Auto-Upgrades Begin on 1 September

Diary entry, and the most actionable item this week. From 1 September, Google Ads will automatically upgrade campaigns using automatically created assets or the campaign-level broad match setting to AI Max for Search campaigns. Google has begun emailing affected advertisers, and one practitioner summed it up as “the great merge continues”.

The upgrade defaults matter:

Campaigns using automatically created assets move to AI Max with search term matching AND text customization enabled by default
Campaigns using campaign-level broad match move to AI Max with search term matching enabled by default

Google says settings will be configured to mirror your legacy setup closely, but text customization arriving switched on is the part to check. That’s the feature which, per Google’s own documentation update last month, generates headlines and descriptions from your landing page content. Any advertiser who deliberately kept copy under human control is about to have that decision reversed by a default.

Audit affected campaigns before the end of August, decide consciously whether each one wants search term matching and text customization, and document the choice. Regulated clients in particular should not discover in October that their ad copy is being written from page content nobody reviewed.

ChatGPT Ads Adds Conversion Bidding and Product Carousels

OpenAI’s advertising platform crossed a threshold this week. Product feed campaigns now support conversion-optimised cost-per-click bidding in beta, letting advertisers optimise towards conversions while still paying per click, with bulk cloning from existing CPC campaigns. Dynamic URL parameters arrived, automatically appending campaign, ad group and ad IDs for attribution. Measurement integrations expanded, including Triple Whale. A multi-product carousel format is in testing, and the platform is preparing to launch in two new markets.

Track the sequence: audiences in July, then exclusions, then multiple advertisers per answer, and now conversion bidding, product feeds, dynamic URLs and third-party measurement. That is not a series of feature announcements, it is a functioning performance channel assembling itself in public over about six weeks.

The window for cheap learning is closing faster than most media plans update. Small structured tests now buy you performance data before the auction fills; waiting for maturity means arriving alongside everyone else with no benchmarks of your own. Pair the platform’s momentum with SE Ranking’s relevance findings above, though, and the sensible posture is enthusiastic testing with tight monitoring of where your ads actually appear.

AI Labels Spread to the Local Pack and Discover

Google’s AI disclosure labels, announced last month for ads created or edited with AI, have started appearing beyond the main sponsored results. They’re now showing on local pack listings and within Google Discover.

The rollout is following exactly the trajectory the original announcement implied, which means agencies still treating AI-usage disclosure as a future admin task are running out of runway. If you followed the advice from a fortnight ago and added an AI-usage field to creative briefs, this changes nothing for you. If you didn’t, local clients are now the priority: the local pack is where a labelled ad sits closest to an unlabelled organic listing, and nobody yet knows whether that comparison helps or hurts click-through. Start collecting your own data on it before somebody asks you a question you cannot answer.

Merchant Center Hits an Outage Mid-Migration

Google confirmed it was investigating an issue affecting Merchant Center this week, without sharing a cause or expected resolution at the time of reporting.

Timing is what makes this more than a routine wobble. Retailers are in the middle of the Content API to Merchant API migration, with the old API’s shutdown deadline having passed on 18 August in Google’s original schedule, and many stores have been reconnecting integrations in recent weeks. An outage lands very differently when half your clients have just changed how their feeds talk to Google.

If a retail client reported product disapprovals, missing items or feed errors around this period, check whether the timing matches the outage window before you spend a day debugging their feed. Log what you see, because platform-side incidents have a habit of being quietly resolved without ever appearing in a status history you can point a client towards.

Microsoft Advertising Lets You Appeal Disapprovals in Bulk

Microsoft has released bulk editing and appealing of disapproved assets, live globally across all Microsoft Advertising accounts.

Anyone who has worked through a mass disapproval one asset at a time will appreciate this more than the announcement’s brevity suggests. Policy sweeps tend to catch dozens of assets simultaneously, and the recovery time has always been dominated by clicking rather than by any actual review. It’s a small operational win that quietly increases how much Microsoft Ads volume a team can manage without extra headcount, which matters given how many accounts treat Microsoft as the channel that gets attention only when something breaks.

Google and Bing Both Redesign the Sponsored Result

Both major engines spent the week experimenting with how paid results look. Google is testing blue hyperlinks inside sponsored ads that lead to the same landing page as the headline, a website-name subheading beneath the Sponsored Results label, and AI-generated descriptions in shopping ads. Search Engine Watch also spotted non-clickable sitelinks appearing beneath ads in AI Mode, a format that displays like sitelinks without behaving like them.

Bing, meanwhile, is testing See more and Visit site buttons on sponsored results, plus something stranger in organic listings: clickable underlined text inside result descriptions that fires a new Bing search rather than sending you to the site.

That Bing test is the one to watch. A search engine turning its own organic snippets into internal navigation keeps users inside Bing for another query cycle, which is the same instinct behind AI Overviews and behind Google’s goto redirect testing earlier in the summer. None of these tests individually changes your work. Collectively they describe an industry redesigning the result page around retention rather than referral, and every point of friction added between a listing and a website eventually shows up in somebody’s click-through rate.

Your Account Average Is Hiding Half the Story

A practitioner piece from PPC Hero worth ten minutes of anyone’s week: a healthy-looking account spending $60,000 a month turned out to be feeding half its budget to junk conversions, entirely hidden behind a respectable blended average.

The mechanism is familiar to anyone who has audited an inherited account. Blended averages combine genuinely profitable segments with genuinely wasteful ones and produce a number that looks acceptable, which is precisely why nobody investigates further. Segmentation is the antidote, and the article walks through which cut exposes the problem fastest.

Given how much of this week’s news pushes towards automation writing its own copy, choosing its own audiences and summarising its own performance, a reminder that account-level numbers conceal as much as they reveal feels well timed. The judgement that automation cannot replace is knowing which segment to look at, and being willing to look when the headline number says everything is fine.

Strategic Direction: Where Search Marketing Is Heading

Two stories this week describe the same shift from opposite ends. Search Console now reports AI impressions to everyone while withholding clicks, and Microsoft Clarity added a Share of Authority metric that measures only the conversations you already appeared in. Both give you more data than you had a month ago; neither closes the gap between visibility and outcome. That gap is where client trust gets won or lost this year, so the practical discipline is to report visibility and revenue as separate lines, resist any metric that quietly redefines the denominator, and keep at least one measurement tool that isn’t sold by the platform being measured.

Meanwhile the platforms keep absorbing the journey. Yelp bookings inside ChatGPT, connected apps inside AI Mode, ads on a quarter of commercial prompts on both major assistants, and a Google homepage testing life without a search button. The through-line is that discovery, comparison and transaction are collapsing into single conversational surfaces owned by two or three companies, with Meta apparently building the index it would need to become the third or fourth. Being the answer increasingly means being in whatever inventory, feed or aggregator the assistant reads, which turns data plumbing into a visibility discipline.

And then there’s cats.txt, which deserves the last word here. An industry rushing to sell tactics for these new surfaces has been accepting evidence that a joke file cleared without effort. The winners over the next year will be the teams who can tell the difference between a mechanism and a coincidence, and who are willing to say “we tested that and it did nothing” out loud to a client.

Key Takeaways

Audit campaigns using automatically created assets or campaign-level broad match before 1 September; AI Max upgrades arrive with text customization switched on by default
Open the Search Console generative AI report on client properties now, and set expectations that it shows impressions only, with no clicks or queries
If you buy ChatGPT ads, monitor placement relevance closely: 14.35% of ads studied were effectively off-topic, and 96.37% of advertisers were not cited in the answer they appeared beside
Revisit your AI crawler policy for Meta specifically, now that its nine billion quarterly requests look like index construction rather than waste
Apply the cats.txt test to every GEO tactic on your roadmap: if crawled, indexed and repeated by an LLM is the whole evidence base, it proves nothing
Report Clarity’s Share of Authority carefully; it covers only queries where you were already cited, not your share of AI search
Update WordPress installations to 7.0.3, which fixes twelve vulnerabilities including a high-severity login screen XSS flaw
Check Business Profile names for repeated bilingual or multi-script versions, which now breach Google’s naming guidelines

Frequently Asked Questions

Does the Search Console generative AI report show AI traffic?

Not yet. It reports impressions, pages, countries, devices and dates for generative AI features across Search and Discover, but it still contains no click or query data. For any traffic question you need your own analytics segmentation alongside it, and you should be explicit with clients about which number came from where.

What do I need to do before AI Max auto-upgrades on 1 September?

Identify every campaign using automatically created assets or the campaign-level broad match setting, because those upgrade automatically. Decide deliberately whether each should keep search term matching and text customization, both of which are enabled by default in the upgrade. Text customization lets Google generate headlines and descriptions from your landing page content, so regulated clients need an explicit decision rather than a default.

Should I advertise on ChatGPT given the relevance problems?

Testing at modest spend still makes sense, because the platform now supports conversion-optimised bidding, product feeds and third-party measurement, and competition remains comparatively thin. Watch placement quality closely though. In SE Ranking’s study around one in seven ads appeared next to a prompt it had little to do with, and buying an ad almost never got the advertiser cited in the answer itself.

Is llms.txt worth implementing?

The honest answer is that nobody has yet shown evidence it works which a file about office cats could not also pass. That does not prove llms.txt is useless, but it does mean the four common proofs (bots crawled it, Google indexed it, an LLM repeated it, ChatGPT endorsed it) tell you nothing about ranking or citation benefit. If you implement it, do so cheaply and without promising clients an outcome.

How worried should we be about AI agents reading hidden text?

Concerned enough to treat it as a live risk in agentic workflows. Forcepoint’s threat model shows hidden page text being absorbed as fact and resurfacing in unrelated recommendations weeks later, demonstrated against several major assistants without any privileged access. If your clients run agents that browse the open web, the pages those agents read are part of your attack surface now.

Conclusion

A week of partial answers. Google gave every property owner a window onto AI visibility and kept the clicks; two studies measured the new ad surfaces and found them lucrative for platforms and unreliable for advertisers; Meta’s crawling turned out to have a purpose after all. None of it demands panic, and all of it demands the same unglamorous response: check the campaigns that auto-upgrade next month, open the new reports and understand exactly what they do and don’t say, update the plugins, and hold every shiny tactic to a standard a joke file about cats could not clear. The teams who keep asking what would look different if this did nothing are the ones whose clients stay confident while everyone else guesses.

Need help adapting your search strategy for the AI era? Contact the Anicca team for expert SEO and PPC guidance.

Table of Contents

SEO & Google Search Updates

AI Search & Visibility

PPC & Paid Media

Wrap-Up