Research · AI search
GEO, measured: three things the market assumes and our data doesn’t support
Generative engine optimisation is now on a lot of marketing service pages. Much of the advice rests on three assumptions. We measure AI answers for a living, so we checked each one against data we had already collected in October 2026.
We build Goppa, which sells this kind of measurement, so we have an interest in the answer. That is why every number here comes with its sample size, its date and the data to check it. No agency or business is named on this page.
1. “AI search is one channel”
The assumption: get into ChatGPT, or “into AI”, and you are in AI search.
businesses that AI assistants recommended in our local searches were named by only one of the three assistants. 250 were named by two, and 47 by all three.
38 searches for dental implants, marketing agencies and web design agencies in a given city, 5 buying questions each, asked on ChatGPT, Google AI Overview (searched from the city) and Perplexity between 21 October 2026 and 30 October 2026.
| Assistant | Answers | Businesses it named | Named by no other assistant |
|---|---|---|---|
| ChatGPT | 190 | 982 | 769 of 982 |
| Google AI Overview | 184 | 562 | 388 of 562 |
| Perplexity | 190 | 777 | 523 of 777 |
Software buyers see the same split. On 100 software buying questions, ChatGPT, Google AI Overview, Google AI Mode, Gemini and Perplexity named the same brand first in 14 of 89 questions where all five answered, and had any brand in common in 37 of 95.
For a business: being named on one assistant says little about the others. Ask for results per assistant, never one blended “AI visibility” score. For an agency: a report that merges assistants hides where the client really shows up and which assistant to work on next.
2. “Optimise your own pages and the AI will pick you”
The assumption: rewrite the client’s site for AI and the answers follow.
Google AI Overviews for our software questions had a YouTube video among their sources, and 54 of 95 a Reddit or Quora thread. ChatGPT cited no video and no forum thread in any of its 100 answers.
Same 100 questions, same day (31 October 2026), United States.
A company’s own site does matter: every assistant cited the own site of a company it named in most of its answers (the first column below). But it is not the only place they read, and each one reads somewhere different. Perplexity cited review sites in 43 of 100 answers; Google’s two answers leaned on videos and forum threads.
| Assistant | Answers | Own site of a named company | YouTube | Reddit or Quora | Review sites (G2, Capterra…) |
|---|---|---|---|---|---|
| ChatGPT | 100 | 96 of 100 | 0 of 100 | 0 of 100 | 17 of 100 |
| Google AI Overview | 95 | 86 of 95 | 84 of 95 | 54 of 95 | 10 of 95 |
| Google AI Mode | 100 | 82 of 100 | 52 of 100 | 50 of 100 | 14 of 100 |
| Gemini | 100 | 88 of 100 | 9 of 100 | 6 of 100 | 23 of 100 |
| Perplexity | 100 | 90 of 100 | 0 of 100 | 1 of 100 | 43 of 100 |
How many answers cited at least one site of each kind. One answer can cite several kinds, so a row does not add up. The full breakdown is in What AI assistants cite when people ask which software to buy.
For a business: where you should appear next (a video review, a forum thread, a review site) depends on the assistant and the category, and it can be measured before anyone writes a line. For an agency: work limited to the client’s own pages covers one part of what the assistants read.
3. “Ask ChatGPT and you’ll see if you’re there”
The assumption: a screenshot of one answer, or an AI chat app asked directly, shows whether a brand is visible.
answers named the brand when a chat assistant answered our 6 buying questions for nike.com from its own memory, without web search. Asked on the assistants people use, with live search, from the United Kingdom: ChatGPT 13 of 18, Perplexity 12 of 18, Google AI Overview 5 of 18.
Live search measured on 12 October 2026: 6 questions × 3 runs on each of the 3 assistants (54 answers). The same questions from memory on 17 October 2026, also 3 runs each; the chat assistant named the brand in 18 of 18 as well when the conversation was about that brand.
The two lists barely overlapped. From memory, the chat assistant listed famous brands. With live search, the assistants read pages and named the shops that sell them. A screenshot is one answer out of many: assistants answer differently each time, which is why every figure on this page is a count of answers, and a provider can change the model behind the same product name without notice. Goppa records the model version reported with each answer, so a change shows up as a break in the series instead of a fake win or loss.
For a business: ask for the questions, the number of answers on each assistant and the dates. For an agency: one answer cannot show change over time; the same questions asked again, assistant by assistant, can.
Before you pay anyone for GEO, ask
- Which assistants will you measure, and will you report each one separately?
- How many questions, and how many answers per question on each assistant?
- Which sources does each assistant cite for my market today, and what will you change there?
- How will you show what changed: the same questions asked again, compared answer by answer?
- What happens to the numbers when an assistant changes its model?
What we did not test
We did not measure whether schema or FAQ markup, statistics or quotes in the text, or llms.txt files change who gets named. We don’t know whether they help, and this page makes no claim about them. Anyone who says they do should be able to show the counts before and after.
See it for your own site
The free check asks ChatGPT the question a buyer would type about your business and shows who it named instead, as a count of answers, with what to fix first. No card.
Method and data
Local searches. 38 searches in 30 cities and 6 countries, 5 buying questions each, asked once on each of the three assistants with live web search (564 answers). The businesses each answer recommended were extracted automatically and spellings of one name counted together, as on our city pages. A failed answer is left out of every count. Software. 100 questions in 10 categories, once on each of five assistants, United States, 31 October 2026; method and dataset on the software study. Memory versus live search. Our internal test (12 October 2026 and 17 October 2026); its individual answers are not published.
Each dataset is one run, in its markets, on its dates: a snapshot, not a trend. Every figure is a count of answers on that day, not a fixed property of an assistant.
How to cite: Goppa Research (2026). GEO, measured: three things the market assumes and our data doesn't support. https://trygoppa.com/study/geo-measured — CC BY 4.0.