The short answer, from our own measurements: the work carries across engines, the result does not. AEO tools that track ChatGPT, Perplexity, Gemini and Google AI Overviews are selling you the checking, not the work, and the checking really is per engine. Each engine fetches from a different index, and the pages it reads for the same question differ so much that a brand can be in most answers on one and almost none on another. On October 12, 2026, across the same 20 buyer questions, Claude named Profound in 60 of its 100 answers and GPT-6 Sol named it in 6 of 100. Same questions, same day.
Which index each engine reads
Everything else in this post follows from one table. The rows are what we have measured ourselves or quoted from a platform's own documentation, as of October 2026; the Perplexity row is the one we have not measured, so it carries only what we know about its crawler.
| Engine | Index it fetches from | Crawler to let through in robots.txt | What we have seen |
|---|---|---|---|
| ChatGPT search, Copilot | Bing | OAI-SearchBot (OpenAI says a site that blocks it "will not be shown in ChatGPT search answers") | Copilot cited our pages 104 times from September 18 to October 15, 0 of them for a question about AEO or GEO |
| Claude | Brave | Claude-SearchBot | On six queries checked by hand on October 13, 45 of the 54 results Claude's search returned were in Brave Search's top 20 |
| Gemini, Google AI Overviews | the normal Google crawl (Google: "no additional requirements to appear in AI Overviews") | Google had 196 of our 210 sitemap pages indexed on October 16 | |
| Perplexity | its own fetch | PerplexityBot | not in our daily readout; we only make sure the crawler is allowed |
The practical consequence showed up on our own site in one week. On October 13 Claude's search could find 7 of the 22 pages we track, and that day we submitted 20 pages to Brave, every page it could not find among them. On October 15 it found all 22. On October 19 the result lists behind Claude's search changed shape, every one of our pages dropped out, and on October 20 it was still 0 of 22. Google's count of the same site did not move. One site, two indexes, two different answers to "can the engine see it at all."
Do AEO tools track ChatGPT, Perplexity, Gemini and Google AI Overviews together?
They do, and they bill by the engine. Peec AI's self-serve plans are $95, $245 and $495 a month for 50, 150 and 350 prompts, each prompt "on three engines of your choice." Profound's free trial runs "50 prompts run daily for 7 days" on ChatGPT, Gemini and Google AI Overviews. Both are from the vendors' own pricing pages as we read them on October 5, 2026. Neither tool changes what any engine reads; a tool shows you the gap per engine and leaves the work to your team.
Our own daily readout does the same with two engines and no product behind it: the 20 frozen buyer questions, once each, to Claude Opus 5 and GPT-6 Sol with web search on, 40 answers a morning since October 15. By October 20 the 7-day pool held 6 mornings and 240 answers: our domain was in the search results of 3 of them, named in 0, cited in 0. Two engines were enough to see that the two paths into an answer are different, which is the point of tracking more than one.
Why the same brand gets opposite results
The engines read different pages, so they name different companies. In the mid-point round of October 12, 100 answers per engine, GPT-6 Sol's most-cited sources were developers.google.com (72 answers), help.openai.com (32) and bing.com (22): platform documentation, with vendor pages only on the 8 questions that ask which provider or what price. Claude's most-cited were a pricing guide at stackmatix.com (23), an agency's roundup at discoveredlabs.com (19) and tryprofound.com (18): third-party comparison posts and vendors' own sites.
The entry route differs too. Three vendors we had not seen in the September 28 baseline entered Claude's answers by October 12, each on the back of its own pages: llmpulse.ai went from 0 to 12 citations, omnius.so from 0 to 12, writesonic.com from 0 to 9. On GPT-6 Sol, the 12 questions that ask how to do something named a vendor in 1 of 60 answers on September 28 and again 1 of 60 on October 12. A page that gets you into Claude's answers can do nothing on the OpenAI side if it is not in Bing's index, and a page in Bing's index does not help on a question GPT-6 Sol answers from Google's documentation.
Even the noise is engine-specific. When we measured our own domain twice on August 16 with nothing changed in between, the largest swing between rounds for the control brands was 3.2 points on Claude for head-category questions and 14.9 points on Gemini for the same questions. A number that moved 10 points on Gemini may have moved on its own.
What is one job, and what is per engine
The on-site work is one job, because every crawler reads the same HTML. The words in the raw page rather than loaded by script; the first sentence under each heading short enough to quote (Claude's retrieval quotes about 150 characters of a page at a time); structured data that says what the company is; a robots.txt that admits OAI-SearchBot, Claude-SearchBot and PerplexityBot. We did the first round of that on our own site in August 2026, and within a week the by-name question on the OpenAI stack went from 0 of 8 correct identities to 6 of 8. It did nothing to the category questions on the two engines we track: 0 of 200 answers named us on September 28, and 0 of 200 on October 12.
Three things are per engine, and none of them is a tool:
- The index check. Is the page in Google, in Bing, in Brave? For our site the answers were yes (Google, 196 of 210 on October 16), yes (Bing, where Copilot had cited us 104 times through October 15) and no (Brave, 0 of 22 on October 20), and the no is the one that decides what Claude can cite.
- The third-party pages. The roundups and pricing guides Claude reads and the documentation GPT-6 Sol reads are different lists; being on one list is being in one engine's answers.
- The reading. Five samples per question per engine, with two control brands you do no work for in the same run, so a 10-point move on one engine is read against that engine's own drift. The method is in how to tell whether AEO worked.
So the honest answer to a founder with one marketer and no budget for a dashboard: do the on-site job once, then ask your buyers' questions yourself on at least two engines that read different indexes, five times each, and write down what each one cited. The cited list is the per-engine work. A tool for each engine would show you the same list for a monthly fee. Our own daily readout is nothing more than that: two engines, one frozen list of 20 questions, no dashboard.
How we checked
Engine behaviour: our own rounds of September 28 and October 12 (20 questions, 5 samples per engine, Claude Opus 5 with web search and GPT-6 Sol through Codex with web search) and the daily readouts from October 15 to 20 (the same 20 questions once each). Index coverage: our Claude retrieval probes of October 13 to 20, Google Search Console on October 16 and Bing Webmaster Tools for September 18 to October 15. Vendor prices: Peec's and Profound's own pricing pages on October 5, 2026. Platform statements: Google's page on AI features and OpenAI's crawler documentation, quoted as fetched on October 18. Perplexity is not in any of our measurements, and the table says so.
Disclosure
ClawWorld is an AEO (Answer Engine Optimization) service for B2B startups based in New York. The measurements in this post are our own, and the places where our own number is zero are left in.