Last Week in AI Search: October 26

The Story

Do you know how ChatGPT searches for your answer? It rewrites your question and never asks exactly what you ask.

No way that could ever go wrong.

When you ask ChatGPT a question, it doesn’t search for your question the way a Google search does. It rewrites your prompt into its own searches, and most people using ChatGPT never see what it searched.

Those searches decide which pages the AI reads. The pages it reads decide who becomes the source. If ChatGPT rewrites every question before it searches, optimizing your site for prompts isn’t the only hurdle to clear for AI visibility. You need to optimize for these back-end searches. And that’s a problem, since ChatGPT is starting to show less of how it searches.

Searchable, an AI-visibility firm, recorded what ChatGPT actually searched. On September 9 and 10, it asked ChatGPT 20 questions about local news and services in each of 100 UK towns and cities. The study captured 9,700 of ChatGPT’s searches.

Not one of those searches repeated the question word for word. The questions averaged 11 words. The searches averaged 7.8.

ChatGPT kept some words and dropped others.

When a question named a town, the town appeared in 98.4% of the searches. The word “local” survived in 38.1%. ChatGPT also added words nobody typed. It put the year into 94.5% of all searches and the month into 75.8%. Of 300 questions that asked about “today” or “right now,” ChatGPT added the exact date to 296 of them.

That isn’t a flaw. A search for “news today” is vague, and a search with the date is precise. ChatGPT’s rewrite in that example makes your query more accurate. But if the searches add words, and still average about three words fewer than the actual prompt, what context are they missing? How does that affect the answer you receive?

ChatGPT also searched more than once. Answers ran 3.2 searches on average, and as many as 35. A question about planning applications averaged 9.5 searches. In a second round, ChatGPT usually searched inside one website, such as a council’s or a local paper’s, using the “site:” operator. That happened in 70.2% of second-round searches, against 11.8% of first-round ones.

Most of what it found, it didn’t use. Only one page in four that the searches returned ended up cited.

So who becomes the source? It isn’t the page that best answers your question. It’s the page that best answers ChatGPT’s version of it.

Other research points the same way. Last week we reported a University of Toronto study of 158,099 AI answers. The researchers asked why ChatGPT cited one page on a website and not another. Signals like FAQ blocks, schema and page speed barely mattered. What mattered was whether ChatGPT’s own searches had turned the page up, which counted for more than ten times as much as the page itself. That mostly comes down to using the words those searches use.

A second study found that ChatGPT adds your city to its searches, using the location your app reports. It does this even when your question never mentions a place.

The searches also change from one run to the next. Seer Interactive ran the same prompt twice, eight minutes apart. The two runs shared zero identical searches.

We know all of this because researchers could capture ChatGPT’s searches from the app. But that view is narrowing.

From September 3 to 16, some of the search data Seer could read vanished from ChatGPT’s stream.

What the record says to do

Write for the searches ChatGPT generates, not just the question people ask.

• Put the year and the place where a search can match them (AI citation). ChatGPT added the year to 94.5% of its searches and kept the town in 98.4%. This week, Sistrix found that the current year in a page’s title was the strongest citation factor it measured. Pages with it had 2.7 times the odds of being cited. The Toronto study found that content still matters, mostly through matching the words of ChatGPT’s rewritten searches. So name the place, the date and the thing plainly. The year has to be true. The record’s rule is to date everything and never bump a date to look fresh.
• Keep the door open for the second round (crawler access). Seven in ten second-round searches searched within a single site. Those searches can only reach sites that let OpenAI’s search bot in. If a site blocks it, ChatGPT won’t use that site in its search answers. At most, it may show a plain link to it. Last week’s Story found 16 of 52 major sites block that bot. Decide deliberately whether to allow each bot.
• Measure more than once (AI citation). If two runs eight minutes apart share no searches, one run is a sample, not a baseline. The record already says to track citation share per engine and to expect it to move. SISTRIX found ChatGPT replaces as much as 74% of its cited sources every week. Before you trust a visibility score, ask how many prompts it tracks and how many times each was run. The IAB calls anything under 50 queries exploratory.

Searchable sells AI-visibility software. Its data covers ChatGPT only, for UK local questions, on two days in September, logged out. Seer’s test was one prompt. The Toronto study is a preprint. Sistrix measured the German-language market. Read the item →

The Follow-Up

Robots.txt decides who gets in, not what happens after, Google says

Last week’s Story found that sites with OpenAI ties let ChatGPT’s search bot in and keep Claude’s out. A second count landed the same day. OpenAEO, an AI-visibility vendor, found 11 of 12 sites allowed OpenAI’s search bot and 4 allowed Claude’s.

This week Google explained the limit of that lock. Robots.txt controls whether a page may be fetched, Google’s Gary Illyes said, and “nothing in it expresses a purpose limitation.” Blocking a bot later has nothing to do with data it already collected.

That’s the fight over Meta’s new business terms. Publishers say a crawling clause lets Meta use their pages for AI. Meta pushed the start date from October 30 to November 20. It hasn’t said it will change the wording.

The Numbers

• Australian vet clinics with 200 or more reviews were recommended 99% of the time by ChatGPT, Gemini and AI Mode. Clinics with fewer than 50 were recommended 37% of the time. (Seenmetric; 360 answers) Read →
• ChatGPT sent 100 sites 0.62% as many visits as Google organic search. Those visitors converted at 13.8%, against 3.02% for organic. Google organic still drove 96.5% of conversions. (Marie Haynes analysis; a second study found a smaller premium) Read →
• Facebook, YouTube and Instagram were the top cited sources in 556 AI answers about seaside holidays. Croatia’s official tourism site wasn’t in the top 50. Spain’s and Greece’s were. (Determ study) Read →

The Rest

• Starting October 30, OpenAI’s models can rewrite advertisers’ ChatGPT ads, including headlines, images and landing-page links, without approval. Advertisers can turn it off. ChatGPT Ads now run in 89 countries. Read →
• Google’s privacy policy now lets Gmail, Drive and Photos data shape ads for users who connect those apps through Personal Intelligence. It adds an exception to a pledge Google had kept in the same words since July 2021. Read →
• In a pilot on about 1% of its network, Cloudflare used signals like page-update recency to help OpenAI’s crawler find pages. The companies say it indexed “hundreds of millions” of pages ChatGPT Search hadn’t found. Read →
• Corrections this week — We merged two stories the wire ran twice (GA4 Direct traffic and DataDome), and corrected the summary of our OpenAI-deals item. All corrections →

---

All 52 items from October 19–25, dated and sourced: the week on the wire · State of Search · Corrections · Archive. Seen something the record should hold? Reply to this email; tips are credited by name.

The full wire

55 items published October 19–25. Every item dated, sourced and linked. An item can carry more than one lane, so lane totals exceed the count.

Research · 23

• Oct 19 — DataDome: Meta’s crawlers drove half of Q3 AI agent traffic — Help Net Security (vendor study)
• Oct 19 — Analysis: GA4 Direct traffic shows little hidden AI activity — Growth Memo (Kevin Indig) (study)
• Oct 19 — ChatGPT, Gemini use different location signals, study finds — arXiv GEO (study)
• Oct 19 — Study prices sellers’ hidden influence on AI shopping agents — arXiv GEO (study)
• Oct 20 — BCA, Mandiri split Indonesia’s AI banking lead — ContentGrip (vendor study)
• Oct 20 — AI Mode cites social sources less than AI Overviews — SE Ranking (vendor study)
• Oct 20 — Cloudways survey: 76% say AI search cut their traffic — Cloudways (vendor study)
• Oct 20 — Study: AI seller agents practice GEO on other AI agents — arXiv GEO (study)
• Oct 21 — Google AI Overviews now reach 2.5 billion people a month — DataReportal (vendor study)
• Oct 21 — Determ study: AI cites social media over tourism sites — Croatia Week (vendor study)
• Oct 21 — B2B brands using influencer marketing get cited more by AI — TopRank Marketing / Ascend2 (vendor study)
• Oct 21 — AI Overviews now cover 88% of commercial search queries — Semrush Blog (vendor study)
• Oct 21 — AI Overviews vs ChatGPT on holiday travel brands — BrightEdge (vendor study)
• Oct 22 — ChatGPT cites official sources over local news, study finds — Press Gazette (vendor study)
• Oct 22 — Sistrix ranks GEO tactics by how often AI engines cite them — Sistrix Blog (vendor study)
• Oct 22 — AI referral traffic converts better, studies find — Marie Haynes
• Oct 22 — ChatGPT search queries shift between identical prompts — Seer Interactive (vendor study)
• Oct 22 — Study: AI search citations shift depending on query language — HackerNoon (single source)
• Oct 22 — Study: AI search tools favor vet clinics with more reviews — Pet Industry News (vendor study)
• Oct 23 — Truelogic: Philippine AI Overview mentions jump 23x — adobo Magazine (vendor study)
• Oct 24 — Study finds safety gaps in Google AI Overviews on first aid — Discover Artificial Intelligence (Springer Nature) (single source)
• Oct 25 — ContextFetch finds bigger sites block AI crawlers more often — ContextFetch (vendor study)
• Oct 25 — Vercatus: AI crawlers read SaaS sites, not travel sites — Vercatus (vendor study)

Law & Money · 8

• Oct 20 — Forbes joins Google AI pilot as referrals drop 45% — Digiday
• Oct 20 — Korean publishers seek AI book-licensing body — Seoul Economic Daily
• Oct 20 — 63 media groups oppose Australia’s AI training opt-out plan — News/Media Alliance (single source)
• Oct 22 — Judge hears fair-use arguments in Concord Music v. Anthropic — Courthouse News Service (single source)
• Oct 22 — News/Media Alliance demands Meta withdraw AI scraping terms — News/Media Alliance (single source)
• Oct 22 — Publishers ask EU, UK to halt Meta AI-scraping terms — Reuters (single source)
• Oct 23 — Bloomsbury’s AI licensing income falls, hurts profit — Bloomsbury Publishing (interim results)
• Oct 23 — Meta delays AI-scraping terms to November 20 — Cryptopolitan (single source)

Tools & Vendors · 7

• Oct 19 — Miracly raises $2M pre-seed for AI search brand monitoring — e27
• Oct 19 — Profound adds Moz data to Pages, next to AI citations — Profound research blog (company announcement)
• Oct 20 — ChatGPT’s dropped UTM tags were a bug, Adobe exec says — Search Engine Journal
• Oct 21 — Peec AI launches Peec Agent, redesigns Actions for AI search — Peec AI blog (company announcement)
• Oct 23 — Reputation launches AI Readiness and AI-ready location pages — Reputation
• Oct 24 — Swap Commerce acquires Vizby AI-visibility platform — Business Wire (Swap Commerce)
• Oct 25 — Forbes blocks PerplexityBot, El País reopens to AI bots — Crawler access panel (first-party data)

Engines · 7

• Oct 19 — ChatGPT Ads adds new ad formats, redesigned conversions page — Search Engine Roundtable
• Oct 19 — DeepSeek opens search engine API, filters GEO manipulation — DeepSeek API documentation
• Oct 20 — Tealium adds ChatGPT Ads pixel and CAPI tracking — Tealium
• Oct 20 — Cloudflare, OpenAI pilot adds pages to ChatGPT Search — Cloudflare (company announcement)
• Oct 20 — ChatGPT Ads will auto-enable creative enhancements Oct. 30 — Search Engine Roundtable
• Oct 21 — ChatGPT Ads goes live in 26 more countries, hits 89 total — OpenAI News (company announcement)
• Oct 23 — Gemini 4 Argon hit 25 trillion tokens in 3 days, Pichai says — PPC Land

Google Search · 7

• Oct 19 — Google tests Preferred Sources button in AI Mode citations — Search Engine Roundtable (single source)
• Oct 20 — Google’s AI Overview label disappears on desktop — Search Engine Roundtable (single source)
• Oct 21 — Google adds AI Overviews opt-out to ‘Control what you share’ guide — Google Search Central (company announcement)
• Oct 21 — Google: robots.txt limits crawling, not AI use — PPC Land (single source)
• Oct 23 — Google’s EEA product grid is gone, Google Shopping leads — BrightEdge (vendor study)
• Oct 23 — Google AI Overviews citation cards get AI-written summaries — Search Engine Roundtable (single source)
• Oct 23 — Google lets Gmail, Photos data shape ads — PPC Land

AEO · 3

• Oct 19 — AI shopping agents strain retailers as purchases quadruple — The Wall Street Journal (vendor study)
• Oct 21 — Mediagene and Infobahn launch AI visibility content marketing business — Mediagene / Infobahn (vendor study)
• Oct 23 — Google: robots.txt limits AI crawling, not AI training use — Google Search Central (YouTube) (company announcement)

Read the complete archive →