AI search optimization is how you get into the answers that Google's AI Overviews and AI Mode, ChatGPT, Gemini and Perplexity write: your page as a source the answer cites, or your brand as a company it names. Each engine can look the question up before it answers, and when it does, it writes from the pages its search returns. So the first job is to be among those pages, or named on them. How a page reads matters after that.
Rank counts for a lot even inside an answer engine. AirOps put 16,851 questions to ChatGPT three times each in a study published in April 2026: a page at the top of ChatGPT's own web search results was cited 58% of the time, and by position 10, 14%. In C-SEO Bench, a controlled lab test, moving a document to the top of what the model read raised its place among the citations far more than any rewriting method did.
Turn AI answers into customers.
See where AI answers recommend someone else to your buyers, why, and the change to make first. Free, no card, about two minutes.
Start freeHow AI search engines choose what to cite
An AI search engine that looks a question up answers in two steps. First it searches, often several times for one question, and gets back a list of pages. Then a language model reads some of those pages and writes the answer, linking to a few of them.
A citation doesn't guarantee a mention, though. In a study published in June 2026, Semrush and Kevin Indig logged 3,981 appearances of a domain in answers from ChatGPT, AI Overviews, AI Mode and Gemini, across 115 prompts in 14 countries. In 61.7% of them, the page was a source and the brand's name never appeared in the answer.
Gemini often answers without any source. In a study Ahrefs published in October 2026, which tracked the same questions for 20 days, 64% of Gemini's answers cited nothing, against under 1% for ChatGPT, AI Overviews and AI Mode. So for Gemini, whether you're named tells you more than whether you're cited.
Where each AI engine gets its pages
Three of these five draw on Google's index, and ChatGPT and Perplexity send crawlers of their own. (Copilot, Claude and other assistants aren't covered here.)
| Engine | Where its sources come from | What has to let it in |
|---|---|---|
| Google AI Overviews | Google's index. A page must be "indexed and eligible to be shown in Google Search with a snippet" | Googlebot, and Search Console's Search generative AI control set to Include |
| Google AI Mode | Google's index, searched several ways for one question | The same as AI Overviews |
| Gemini app | The Google Search index, which Google says it uses to ground Gemini's answers | Googlebot, and no Google-Extended block in robots.txt |
| ChatGPT | Pages collected by OpenAI's search crawler, plus results from a third-party search provider | OAI-SearchBot, and ChatGPT-User, which fetches pages while it answers |
| Perplexity | Its own index, which Perplexity says covers "hundreds of billions of webpages" | PerplexityBot, and Perplexity-User, which fetches pages while it answers |
Google-Extended isn't a crawler. It's a name you can block in robots.txt to keep your pages out of Gemini's training and grounding. Google says it doesn't affect inclusion in Google Search, so a block doesn't take you out of AI Overviews or AI Mode.
For the two fetchers that open pages mid-answer, ChatGPT-User and Perplexity-User, the control that counts is your CDN or firewall. They fetch on a person's behalf, and their vendors say that can put them outside robots.txt. The AI crawlers page links each user agent to the IP ranges its vendor publishes, and OpenAI's crawlers are covered in more detail in How to rank in ChatGPT.
How to rank in AI search, in order
The first two steps decide whether your own pages can be cited. The third gets you named on other people's pages, and it's where to start while your pages don't yet rank for your buyers' questions. The fourth helps a page the engines already read.
1. Let each engine's crawler read the page
A block here hides you from that engine whatever you do later, and it takes minutes to rule out.
robots.txt. A rule against OAI-SearchBot or PerplexityBot can keep your pages out of that engine's search, and a rule against Google-Extended keeps them out of Gemini's grounding. GPTBot collects training data for OpenAI, and blocking it has no documented effect on ChatGPT's search. A free crawler check shows which AI crawlers your robots.txt turns away from your homepage. Blocks like this seem to be uncommon: when I scanned 500 YC startups in October 2026, 3 of the 471 sites I could check (0.6%, 0.2–1.9%) shut out an AI answer crawler in robots.txt, and one of the three did it with a Disallow: / for every crawler.
Your CDN or firewall. A CDN's bot settings act before a request reaches your server, so a refusal there leaves no line in your own logs, and robots.txt has no say in it. On Cloudflare, read the AI crawler settings it introduced on 15 September 2026.
The HTML your server sends. When ChatGPT or Perplexity opens a page in the middle of an answer, it reads the HTML as sent, judging by published rendering tests from 2026, so text your scripts add later may not reach it. In the same YC scan, 25 of the 453 homepages I could read (5.5%, 3.8–8.0%) had no text until their scripts ran, after a re-check of the borderline ones.
To see what the crawlers got, read your server logs for OAI-SearchBot and PerplexityBot requests and their status codes. Anyone can put those names in a request, so check a hit's IP against the vendor's published ranges before acting on it. For Google, the URL Inspection tool in Search Console shows whether Google could fetch a page and whether the URL is on Google.
2. Rank in the searches each engine runs
For the three engines on Google's index, this is mostly Google SEO. The page has to rank for the question, or for one of the follow-up searches an engine runs on the question's behalf.
Of the 112 Google searches with an AI Overview in a study I ran in October 2026 on 40 uptime-monitoring questions, Better Stack's pages were on page 1 for 9, and none of those 9 Overviews cited them. UptimeRobot's site was on page 1 for 36, and 29 of those Overviews cited a page on its site, against 15 of the other 76. Why AI search keeps naming UptimeRobot has the rest.
For ChatGPT, rank inside its own search results counted in the AirOps study, but that rank is hard to see from outside ChatGPT, and studies of the usual stand-ins disagree. When Seer Interactive checked 100 questions in February 2025, Bing's top 20 held 87% of the pages ChatGPT cited. In Ahrefs' 2025 study of 15,000 prompts, about 8% of ChatGPT's citations ranked in Bing's top 10 for the same question, and about 8% in Google's top 10. An October 2026 preprint, built on 80 questions, found that Google Search and OpenAI's own search API, combined, returned only about a quarter of the exact URLs ChatGPT cited.
So rank in Google, which feeds three of the five engines, but don't read Google or Bing rankings as a forecast for ChatGPT. A more direct sign that ChatGPT can find a page is OAI-SearchBot fetching it.
Perplexity searches its own index, and I haven't found a study of how the Perplexity app's citations line up with Google's or Bing's results. PerplexityBot requests in your logs are a sign it has your pages.
3. Get onto the pages the answers cite
On questions that don't name a brand, most of what AI answers cite is other people's pages. Momentic tracked 21 brands from 1 September to 18 October 2026. On category questions, the brands' own websites made up 3.0% of 410,143 citations on AI platforms, and on questions that named the brand, 31.6%. Momentic notes that levels may differ in other industries.
Lists make up a large share. In Ahrefs' study of ChatGPT's answers to 750 top-of-funnel prompts, published in December 2025, "best X" blog lists were 43.8% of the source links, and of 3,000 cited lists it checked, 35% sat on low-authority domains. The list that matters in your category may be on a small site you could ask to include you.
The engines don't lean on the same kinds of page. Two 2026 studies counted how often three of them cited YouTube and Reddit:
| Cited in the answer | AI Overviews | AI Mode | ChatGPT | Study |
|---|---|---|---|---|
| YouTube | 36.5% | 35.6% | 2.1% | Foundation: 5,000 B2B software prompts, 13 weekly runs, June–September 2026 |
| 16% | 7% | 1% | TJ Digital: 193,507 answers to 4,167 buyer questions, August–October 2026 |
TJ Digital's YouTube shares were lower than Foundation's, at 23% for AI Overviews and 9% for AI Mode, on a different set of questions, so hold the exact figures loosely. Either way, a video that answers a buyer's question has a better chance of being cited on Google's surfaces than in ChatGPT.
Being mentioned elsewhere goes with being named. Across 10,000 US brands, Ahrefs found in November 2026 that the brands named most in AI Overviews, AI Mode and ChatGPT tended to be the ones most named in Reddit threads that rank on Google (a rank correlation of 0.75), then on web pages (0.70), in news articles (0.67) and on YouTube (0.64). Size accounts for part of that: among brands of similar size, the web-page figure fell to 0.34–0.54. It's a correlation across brands, so it doesn't show what one new mention does.
What to do: for each buyer question, collect the pages cited across several runs on each engine. Sort them into lists you could join, threads you could take part in, review sites, and questions a video of yours could answer. Start with the pages cited most recently, because sources rotate. Why ChatGPT recommends your competitor has a PowerShell check that counts your name and a competitor's on each cited page.
4. Answer the question on your page, and name your product
This step's evidence is the thinnest of the four. In the AirOps study, pages whose best heading matched the question almost exactly were cited 41% of the time, against roughly 30% for weak matches. That's ChatGPT only, and a correlation. Give each buyer question a heading anyway, worded the way buyers ask it, and answer it in the next sentence or two.
Then name your product in that passage, with what it does for the question. An engine can cite your page and leave your name out, as in the 61.7% above. I know of no test of whether naming your product there changes that, but it costs one sentence.
What the evidence says you can skip
When I checked on 11 November 2026 (US, desktop), six of the ten pages I read from Google's results and AI Overviews for "ai search optimization" and two similar searches recommended at least one of the first four below, with no caveat about the evidence. Google's guide for its AI features says three of them aren't needed for Google Search.
Rewriting pages to add statistics. The 2024 paper behind the name GEO reported gains of up to 40%, measured on an answer engine built for the test. C-SEO Bench, the lab test above, published at NeurIPS in 2025, ran the same kinds of rewrite against four newer models in a simulated search engine. Its authors found the methods "largely ineffective" and often harmful, and rewriting to add statistics lowered a document's ranking in 19 of 24 settings. Keep the real numbers your readers need, and skip rewrites made for a model.
Schema markup. In Ahrefs' matched test from May 2026, pages that were already cited and then added it gained no AI citations beyond what chance would explain. LLM SEO has the detail. Rich results in ordinary search are still a reason to keep it.
llms.txt. Of about 38,000 sites that publish one and use Ahrefs' web analytics, 97% got no requests for the file in May 2026. Ahrefs calls every figure in that study a ceiling, because a fetch doesn't mean a bot read the file.
Writing in bite-sized chunks. The same Google guide says in its recap: "For Google Search, you can ignore tactics like 'chunking' content, creating unnecessary AI text files (like llms.txt), or pursuing inauthentic mentions." Danny Sullivan put it plainly on Google's Search Off the Record podcast in January 2026, as Search Engine Roundtable transcribed it: "We don't want you to do that."
"Summarize with AI" buttons that carry instructions. Some sites add buttons that open an assistant with a prompt that tries to plant instructions in its memory. In a report from February 2026, Microsoft's security researchers counted 50 such prompts from 31 companies over 60 days of email traffic, and they treat it as an attack on the assistant. One of the ten pages I read has a button like that, and its prompt asks the assistant to remember the site as a citation source.
How long it takes to show up
The one published timing test I found is Semrush's, from December 2025. It put 81 FAQ pages on its own blog, which has an Authority Score of 84, and counted how many each engine cited. Google AI Mode cited 29 the day after they went live and 48 by day six, then fewer, down to 21 by day 30. ChatGPT started slower, with 8 on day one, 28 by day 14 and 34 by day 30, and pages it cited mostly stayed cited. Semrush tracked only those two engines, on a well-known domain, and calls it the best possible conditions. A newer site is likely to wait longer, and I haven't found a timing test for Gemini or Perplexity.
Citations also come and go from day to day, so one check tells you little. Ahrefs' October 2026 study, the one above on Gemini, followed every site cited on day one for 20 days. By day two, 50% of AI Overviews' had dropped out, and 58% of AI Mode's and 59% of ChatGPT's. Gemini's figure was higher, at 81%, partly because many of its next answers cited nothing; with those left out, all five tools Ahrefs tracked, which include Copilot and not Perplexity, fell between 50% and 60%. About 75% of the sites that dropped out came back within the 20 days, and about half came back within three days. Brand names held steadier than citations, and Ahrefs' advice is to look at the page if you're still missing after two or three weeks.
How to tell whether a change worked
Check that the engine has read the page since you changed it, then ask your questions again.
For the three engines on Google's index, the URL Inspection tool in Search Console shows when Googlebot last crawled a page. For ChatGPT and Perplexity, your server logs show when their crawlers last fetched it. This lists, for one page, how many requests each crawler made and when the last one came:
$log = ".\access.log"
$page = "/pricing"
foreach ($bot in "OAI-SearchBot", "ChatGPT-User", "PerplexityBot", "Perplexity-User") {
$hits = @(Select-String -Path $log -Pattern $bot -SimpleMatch |
Where-Object { $_.Line -match ('"GET ' + [regex]::Escape($page) + '[ ?]') })
$last = if ($hits.Count) { ($hits[-1].Line -split '[\[\]]')[1] } else { "none" }
"{0,-16} {1,5} requests, last {2}" -f $bot, $hits.Count, $last
}
It assumes the common log layout, with the time in square brackets and the request in quotes, and a log written in time order. "none" can mean the crawler hasn't been, or that your log uses another layout. It also trusts the user agent, which anyone can send, so check the IPs of the hits that matter.
Once the page has been read again, repeat the measurement you made before the change, with the same questions and the same number of runs on each engine. Then compare the two rates at which you're named, each with its range. How to measure AI visibility explains the counting, prompt tracking covers which questions to ask, and the best AI visibility tools compares nine trackers that do the asking for you. Google adds a count of its own: Search Console's generative AI report tallies impressions, each one a link to your site shown in one of Google's AI features, but holds nothing on which brands the answer named (what an AI impression counts).
Where Shruwd fits
Shruwd, the tool we are building, measures the answers and the crawler visits described above. It puts your buyers' questions to all five engines and shows how often each one names you and the competitors you track, with a range on every rate. Google's answers come from its results pages, ChatGPT's and Gemini's from their apps, and Perplexity's from its Sonar API, which can answer differently from the app people use. Each is a proxy for what one buyer sees. Your server logs show the crawler side: Shruwd verifies each AI crawler visit against the IP ranges its vendor publishes, and flags a page that keeps refusing a verified crawler. When a competitor is named far more often than you on a question, its findings can name the lists, review sites and threads the answers read, with the day each was last cited. On paid plans it measures again after you mark a fix applied.
Frequently asked questions about AI search optimization
What is AI search optimization?
It's the practice of getting into the answers AI search engines write, either as a source the answer cites or as a company it names. The engines include Google's AI Overviews and AI Mode, ChatGPT, Gemini and Perplexity. Each of them can search the web before it answers, so most of the work is SEO. Other names for it are GEO, AEO and LLM SEO.
How do I optimize my website for AI search?
Start with where each engine gets its pages. AI Overviews, AI Mode and Gemini use Google's index, so let Googlebot in and rank in Google for your buyers' questions and the follow-up searches behind them. Keep Search Console's Search generative AI control on Include for AI Overviews and AI Mode, and leave Google-Extended unblocked for Gemini. ChatGPT and Perplexity crawl for themselves, so make sure neither robots.txt nor your firewall turns away OAI-SearchBot or PerplexityBot, and that each page's text is in its HTML. Then get onto the lists, threads and videos the answers cite, and give each question a heading on your page, answered right below, with your product named. Before you judge a change, check that the engine has read the page again.
Is AI search optimization different from SEO?
Google's guide calls optimizing for its own AI features "still SEO". The differences come from the other engines. ChatGPT and Perplexity send crawlers of their own, Google rankings don't predict ChatGPT's citations well, and in two 2026 studies ChatGPT cited YouTube and Reddit far less often than Google's engines did. GEO vs SEO covers the rest.
How do I rank in Google's AI Overviews?
Make sure Google indexes the page and can show it with a snippet, and that Search Console's Search generative AI control is still on Include, its default. Excluding a site there takes its links and content out of AI Overviews and AI Mode. Then rank for the question or the follow-up searches behind it, and look at what the Overviews for your questions cite. In Foundation's 2026 study of 5,000 B2B software prompts, 36.5% of AI Overview answers cited YouTube, so a video that answers the question is worth considering.
How do I rank in Perplexity?
Let PerplexityBot crawl your site, because Perplexity answers from its own index, and check that your CDN or firewall doesn't refuse Perplexity-User, which fetches pages while Perplexity writes an answer. Then find the pages it cites when you ask your buyers' questions, and work toward being named on them.
How long does AI search optimization take?
On a strong domain, days to weeks for a new page, going by the one published test I found. Semrush put 81 new pages on its own blog and reported in December 2025 that Google AI Mode cited 29 of them the day after they went live, and ChatGPT took about two weeks to cite a similar number. It tracked only those two engines, and a newer site is likely to wait longer. Answers also change from day to day, so check that the engine has read the page again, then give a change two or three weeks before you judge it.
Do I need to optimize for each AI engine separately?
Mostly not. Three of the five engines draw on Google's index, so Google SEO covers much of the ground. ChatGPT and Perplexity add their own crawlers to let in, and the engines differ in what they cite: in Foundation's 2026 study of B2B software prompts, YouTube showed up in about a third of AI Overview and AI Mode answers and in about 2% of ChatGPT's. Ask your questions on each engine and see which pages each one uses.
