Google Search Console's generative AI performance report has had an unresolved logging bug since August 13, 2026, that undercounts the AI citation impressions it shows for pages appearing in AI Overviews and AI Mode. In this episode of The GEO Show, Paris Childress covers a week of AI search developments, from SimilarWeb's data on ads appearing in roughly one in four ChatGPT responses to Apple's sudden 194% expansion of Applebot's crawl infrastructure, Ahrefs' new AI-adjusted volume metric, and research showing brands can be miscategorized by AI despite strong SEO authority.
Generative Engine Optimization has a measurement problem, and this week's news makes that hard to miss. Google's own AI citation data has a bug, Ahrefs is trying to build a search-volume equivalent for AI prompts that may never exist, and new brand research shows large language models can know a company well while still filing it under the wrong category. Paris Childress runs through seven developments from his SEO and GEO briefing: AI search is fast becoming its own ecosystem, with its own analytics, ad infrastructure, and crawlers, but the tooling is still catching up to the reality.
Google Search Console's generative AI performance report has had a logging error since August 13, 2026, that has artificially reduced the impressions it shows, and Google has not yet resolved it. The report is a new addition to Search Console that surfaces citation appearances, what Google calls impressions, at the page level for AI Overviews and AI Mode. Paris flags this as a maturity problem, not a reason to ignore the tool: first-party GEO data from Google is valuable, but data from the affected window should not be trusted until Google confirms a fix.
More than marketers might assume. SimilarWeb has launched a competitive intelligence dataset that tracks ad placements across ChatGPT, Google AI Mode, and AI Overviews using real user panel data, and the early numbers are striking: 26% of ChatGPT responses now carry sponsored ads, and nearly 30% of ad-eligible AI Mode queries show ads.
The attribution infrastructure is catching up just as fast. OpenAI has switched on Automatic Advanced Matching, or AAM, for existing ChatGPT advertisers: by default its ad pixel can detect customer information in forms, normalize it with SHA-256 hashing in the browser, and use that hashed data for conversion matching when click IDs are unavailable. Advertisers can opt out. Paris sees this as OpenAI rapidly building the same attribution stack that Google and Meta have run for years, starting with the pixel.
Apple has expanded Applebot's published crawl infrastructure by 4,656 IP addresses across 21 new ranges, a 194% increase, and offered no explanation. That silence is the interesting part. Applebot has always had a much smaller footprint than Googlebot or the newer AI crawlers from OpenAI, so a jump this size is a real infrastructure signal, not routine scaling. Paris's read is that it points toward Apple building its own AI search capability, possibly one running partly on-device across iPhone, iPad, and Mac, with a public product landing later this year or next.
Ahrefs thinks it is worth trying. Its new AI-adjusted volume metric, part of the Brand Radar feature, estimates platform-specific AI demand for a given keyword using a ratio-based method rather than direct measurement of AI impressions:
Ahrefs itself says this is not actual AI impressions or an AI search total addressable market, just an approximation. Paris is skeptical it will ever become more:
I do think that they are still very far off from being able to forecast and report on prompt volumes. I just don't think that's ever going to exist.
Part of the problem: the underlying signal is disappearing. Google's own AI Mode and AI Max are already removing keyword displays, and Paris doubts OpenAI will expose prompt volume through its ad platform either. Keyword volume is sunsetting, and nothing equivalent is reliably taking its place.
New research from Fractal, covered by Search Engine Land, analyzed 4,320 AI responses and 8,500 unique brand references across multiple models. It found that some strong, traditional search brands barely appeared for categories they expected to compete in, while other brands materially overperformed relative to their SEO authority. The mechanism matters more than the headline: a model can hold accurate, detailed information about a company and still fail to associate it with the category it wants to own.
Paris calls this a distinct GEO problem, category association, separate from factual accuracy. Fixing it means reinforcing category position through comparisons, customer proof, partner references, reviews, and earned media, not just feeding LLMs more accurate facts in isolation.
A related finding shows how ChatGPT itself retrieves information. According to Resoneo's August update, one ChatGPT mode drew 74.7% of its results from OpenAI's own retrieval system, while another drew roughly 75% from Google-derived results, and site-specific search operators rose sharply at higher reasoning levels. Paris's takeaway is blunt: ranking in ChatGPT is becoming a nearly meaningless single KPI, because different modes pull from different webs entirely. The goal is to get the right brand passages and claims surfaced, not to rank a page.
These seven stories are pieces of the same shift: search is splitting into multiple retrieval systems, each with its own crawler, ad stack, and blind spots. The brands that adapt fastest will treat GEO data as a work in progress, not a finished dashboard.
Hello, and welcome back to The GEO Show. I'm your host, Paris Childress, and we're recording on Friday, August 21. This is episode two. I've got a crop of new stories I want to run through today from my SEO and GEO briefing, and I'll give you some of my thoughts. Let's start with the first story: Google's new GEO reporting currently has a data bug.
Google says a logging error has artificially reduced impressions in Google Search Console's generative AI performance report since August 13, and it remains unresolved. This is about one week old at the time of this recording, and it's very interesting because this is a new part of Google Search Console, the generative AI performance report.
I think it has just been rolled out globally, and Google is still struggling to get this right. So what it tells me is that first-party GEO data is extremely valuable, but it's not yet mature enough to trust blindly. That story came from Search Engine Land.
For those of you who are not aware, Google Search Console now shows the citations of your pages. You go into the generative AI performance report and you can see, at the page level, the number of what they call impressions, which effectively means citation appearances in AI Overviews and AI Mode. But don't trust the data from the period of August 13 until, well, at least I think until today.
Story number two: SimilarWeb has launched competitive intelligence for ads inside AI answers. SimilarWeb's new AI ads dataset tracks placements across ChatGPT, Google AI Mode, and AI Overviews using real user panel data. SimilarWeb says that 26% of ChatGPT responses currently carry sponsored ads, and nearly 30% of ad-eligible AI Mode queries now show ads.
So we're talking about a sizable chunk of responses from ChatGPT and Google's AI Mode. 26% of responses from ChatGPT and 30% from AI Mode are now, in the case of ChatGPT, carrying sponsored ads, and in the case of Google AI Mode, eligible to show ads. This is really proving that ads have already started to make a major dent in the inventory of AI chat. It's also interesting that paid AI search is becoming measurable enough to operate as a real media channel.
So that's the beginning of a whole new paid advertising channel that's opening up right now, and it's going to be an interesting story to follow. Moving on to the next story: OpenAI reported in its help center that it has switched on a more aggressive conversion matching system for existing advertisers. It has enabled something called Automatic Advanced Matching, or AAM, for existing ChatGPT ads.
By default, the pixel can detect supported customer information in forms, normalize it using SHA-256 hashing in the browser, and use it when click IDs are unavailable. Advertisers can, of course, opt out of the pixel. That's a familiar term we've heard before: the pixel.
I know we've talked about Facebook pixels hundreds of times in our agency, and now ChatGPT and OpenAI ads have their own pixel. They are rapidly acquiring the attribution infrastructure of Google and Meta, starting with the pixel. That is a real sign of the maturity of OpenAI's ad platform.
The next story is coming to us from PPC land, and it's about Apple. Apple dramatically expanded Applebot's published crawl infrastructure by adding 4,656 IP addresses across 21 new ranges, a 194% increase. Apple did not explain that change in any way.
But this is a very interesting and meaningful infrastructure signal, because it is the first time I have really even heard or thought about Applebot. Of course, we've always known about Googlebot and now OpenAI's bots. But Apple has Applebot, and Applebot is dramatically ramping up its crawl infrastructure. So what that could mean is anybody's guess.
Are they going to launch an Apple AI search engine or an LLM of their own? Who knows. Is that going to be something that might live on the edge, on devices like iPhones, iPads, and Macs? I think most likely, yes, actually. That's probably the reason why Apple is ramping up its investment in its crawl infrastructure.
I do think they're headed toward on-device AI, and that is going to be a major play, probably coming later this year or next. Moving on to the next story, and that is coming from Ahrefs. Ahrefs introduced an AI-adjusted volume metric for prompt demand.
So what does that mean, AI-adjusted volume? Ahrefs has always measured keyword volume. Now this is AI-adjusted volume for prompts. I think what they're attempting to do here is build a metric for AI prompts that is similar to the keyword volume metric.
It's confirmed that Brand Radar, their AI feature, now estimates platform-specific AI demand by taking a Google keyword volume and applying ratios derived from Ahrefs' aggregated AI referral traffic relative to Google organic traffic. Ahrefs explicitly says the metric is not actual AI impressions or an AI search TAM, but that they're trying to get closer to keyword volume for AI prompts. Let's unpack this method.
They're taking keyword volume, which they have good, accurate data for, and applying ratios derived from Ahrefs' aggregated AI referral traffic relative to Google organic traffic. So let's just say, for the sake of argument, their Google organic traffic represents 50% of all traffic coming into Ahrefs, and the AI referral traffic is 5%.
They're taking that ratio, a one-to-ten ratio from five to fifty, and saying that if the keyword volume sending that organic traffic is 50, then the AI traffic must be worth five. The AI-adjusted volume in this case is that 5% number, or the volume of that 5%.
I think that's a pretty interesting approach. I do still think, though, that Ahrefs, and any other classic keyword research tool or SEO tool, are still very far off from being able to forecast and report on prompt volumes.
I just don't think that's ever going to exist. I don't think that even OpenAI, through its ad platform, is going to surface that anytime soon. And we already see that Google's AI Mode and AI Max are already removing the keywords.
So they also don't want you to start, or to continue, researching keywords and keyword volumes as you transition into AI Max in Google Ads. Keyword volume is clearly sunsetting here, and coming in its place will not be prompt volume.
That is not great for attribution and measurement, but that is the real world we're heading toward. Let's go on to the next story. There's new brand research highlighting that AI has a categorization problem. A company called Fractal analyzed, for Search Engine Land, 4,320 responses and 8,500 unique brand references across multiple models.
It found that strong traditional search brands barely appeared for categories they expected to compete in, while other brands materially overperformed their SEO authority. Bottom line: a model can know a company extremely well, yet fail to associate the company with the category it wants to own.
So what does this mean, actually? It's really interesting, because AI can still have a lot of accurate information about your brand but still miscategorize you. That is a GEO problem: category association.
A lot of the SEO and GEO work we should be doing is not only informing LLM crawlers about our brand with accurate claims, but also reinforcing our positioning through comparisons, customer proof, partner references, reviews, and earned media. By comparing our brand to other brands in the category, we are reinforcing our category position and hopefully giving AI crawlers the right categorization information.
Now we're moving on to the next story. This will be our last story of the day: ChatGPT's different modes increasingly search different webs. That's interesting.
This is from Search Engine Land. According to Resoneo's August update, Freethink drew 74.7% of results from OpenAI's own retrieval system, while paid thinking drew roughly 75% from Google-derived results. Paid thinking also narrowed its source set from July to August, while use of site-colon searches increased sharply at higher reasoning levels. This is observational vendor research; it's not OpenAI documentation.
The interpretation here is that ranking in ChatGPT is becoming an almost meaningless single KPI. That's something I think we've always suspected as we move from SEO into GEO: this is no longer a game of ranking. We are not trying to rank pages; we are trying to surface brand passages that are embedded in pages. We're not trying to rank pages, because it's essentially a meaningless exercise. There's too much volatility.
That's a wrap for this episode number two. I hope you all enjoyed it, and I'm looking forward to seeing you next time. So long.
