Short answer: Most of the time ChatGPT does not search at all: Ahrefs analyzed 66.6 million answers and only 14.4% triggered a web search, and when it does search it increasingly adds "official", "price" and "2026". SME websites should state prices, specs and service scope in plain text so AI can find the "official" answer.
On October 12, Ahrefs published "How ChatGPT Does Query Fan-Out", analyzing 66.6 million logged-out ChatGPT (GPT-5.6 Luna) answers and about 15 million fan-out queries from August and September 2026, meaning the search strings AI generates by splitting one user prompt into several. The number that stops you: only 14.4% of answers searched the web; the other 85% came straight from model memory. In one sentence: to be recommended by AI you first need to be a brand the model remembers, then the official source when it does search.
Why does this data matter now? Over the last 18 months, the sources AI answers cite have been reshuffled. The same study shows Reddit fell from roughly 1 in 5 ChatGPT answers in March to 1 in 43 by late September (−88%), and Wikipedia from 1 in 7 to 1 in 35 (−80%). Official sites took their place: in a test of 20,000 US prompts, 84.4% of fact questions cited at least one official source. Google's documentation likewise says its AI features use the same index and quality systems as regular search. Platforms now weigh who said it more than who wrote the most.
Tools that measure this are crowding into the same lane. Ahrefs uses its own Brand Radar citation data, Semrush has an AI visibility toolkit and teaches dashboard building, and startups such as Profound and Scrunch mostly sell enterprise monitoring. When everyone sells "how often AI mentions you", the data quickly becomes table stakes and the only difference left is sample size and interpretation.
The impact on small businesses is direct: when a customer asks AI "how much does a website cost in Taichung" or "how long is this brand's warranty", whether AI lands on your site depends on whether you wrote the answer as text it can read. Below we break down the data, then a playbook that needs no SaaS subscription.
The details: what words AI searches for
Bottom line: fan-out queries increasingly look like a shopper comparing prices, adding year, official, price and reviews.
| Metric | Data | What it means |
|---|---|---|
| Share of answers that search | 14.4% | Most answers rely on model memory; brand awareness comes first |
| Fan-out queries per search | About 1.6 | AI reads results for only a few queries; only top positions get seen |
| English queries with "official" | 10.1% in August → 13.1% in September | AI actively looks for the official word |
| Queries with "price" | 6.0% → 8.9% (about +47% relative) | Pricing pages become search targets |
| Queries with "2026" | 17.6% → 22.9% | Undated old pages lose out |
| Queries using site: | 6.8%, of which 75.9% target brand sites | AI goes straight into your site for answers |
Winners are brand sites that state facts clearly and original research; losers are sites that only summarize others. Comparison questions cite official sources less often (68.3%), so "A vs B" content still leaves room for third-party testing.
What three kinds of readers should do now
Bottom line: owners add official facts, marketers rewrite page openings, developers make sure crawlers get in.
- Business owners: list the 10 facts customers ask most (price range, warranty, service area, lead time) and make sure your site answers each in one sentence.
- Marketing/SEO practitioners: add the year and "official" meaning to key page titles and first paragraphs, e.g. "2026 pricing (official)"; keep comparison posts but back them with your own test data.
- Developers/agencies: check robots.txt is not blocking OAI-SearchBot, and do not put prices and specs only in images, PDFs or behind logins.
AI visibility tools compared
Bottom line: the tools suit brands with marketing budgets; SMEs can start with manual testing.
| Tool | What it does | Price tier | Best for |
|---|---|---|---|
| Ahrefs Brand Radar | Tracks brand mentions and citations in ChatGPT, AI Overviews and more | Add-on to an Ahrefs subscription | SEO teams already on Ahrefs |
| Semrush AI visibility toolkit | AI mention tracking and dashboards | Add-on to Semrush plans | Agencies, multi-brand teams |
| Profound / Scrunch | Enterprise AI answer monitoring and optimization advice | Mostly enterprise quotes | Mid-to-large brands |
| Manual testing + server logs | Ask AI a fixed question set monthly and compare with crawler logs | Nearly free, about 2 hours a month | SMEs |
What the study won't tell you
Bottom line: this is two months of one model in the US English market and cannot be applied directly to Traditional Chinese search.
- No Chinese sample: the "official" figure covers English queries only. In our own September crawler logs, ChatGPT-User fetched pages 501 times, 499 of them English and only 2 Chinese; Traditional Chinese sites may be searched even less.
- 85% of answers never search: most recommendations come from model memory, so even a perfect site needs long-term brand mentions and press coverage to get into that memory.
- A model update resets everything: the author admits the results reflect a single model in August–September, and the next version may search very differently.
A no-SaaS alternative for SMEs
Bottom line: one "official facts page" plus a 20-question monthly test shows a difference within a month.
- Official facts page: one page listing price ranges, plan contents, warranty, service area, contact details and last-updated date, all as plain HTML text.
- Bilingual: at least translate the facts page and main service pages into English, matching AI's current preference for English pages.
- Structured data: add Organization, Product or Service and FAQPage so each field has a clear meaning.
- 20-question monthly test: ask ChatGPT, Gemini and Perplexity questions that include "official", "price" and "2026", and record whether they cite you.
- Read server logs: count which pages OAI-SearchBot and ChatGPT-User fetched, and rewrite the ones they skip first.
FAQ
Bottom line: SMEs mostly ask whether AI will find them, what to write and whether to buy tools.
Does ChatGPT search the web for every answer?
No. Ahrefs found only 14.4% of answers triggered a search; the rest came from model memory.
How should a website be written to get cited by AI?
State facts like price, specs and warranty as one-sentence plain-text answers, with the year and an update date.
Do SMEs need an AI visibility tool?
Not at first. A 20-question monthly manual test plus server logs gives a good picture.
Does a Chinese-language site need an English version?
At least for the facts page and main service pages; AI crawlers currently fetch English pages far more often.
Our take
Bottom line: we think the rise of "official" in AI queries is a bigger opportunity for SMEs than for big brands, because a small company's website is usually its only official source.
The mainstream reading is that AI only cites big brands and small firms have no chance. Our call is the opposite: big-brand information is scattered across news, resellers and review sites, so AI must pick among many sources; for a local small business, the website is the only place its prices and service scope exist online, and if it is written clearly it becomes the only answer AI can cite. The real risk is that many SMEs keep prices only in LINE chats or quote PDFs, where AI cannot see them. Another unpopular call: AI visibility dashboards will become a free bundled feature of SEO tools within 18 months, so paying a premium now is poor value.
The lesson for ScriptWalker (a Taiwan-based Laravel/Flutter custom development studio): package "official facts page + bilingual pages + structured data + crawler log report" as a fixed-price GEO foundation service, which fits SME budgets better than selling monitoring subscriptions.
Sources
- Primary: Ahrefs: How ChatGPT Does Query Fan-Out (15 Million Queries Studied) (2026-10-12)
- Primary: OpenAI: Overview of OpenAI Crawlers (OAI-SearchBot, ChatGPT-User)
- Primary: Google Search Central: AI features and your website
- Third party: Search Engine Roundtable: Daily Search Forum Recap (2026-10-13)
- Third party: Lumar: Which Sections of Your Content Are Actually Citable?
- Third party: Semrush: How to build an AI visibility dashboard