How to get cited by AI answer engines: five steps you can verify
Getting cited by AI answer engines takes five steps in order: allow the citation crawlers, get the page discovered rather than merely reachable, write a short passage an engine can lift, name a source behind every number, and measure which engines cite the page. Each step below carries the check that proves the step is done.
What does it take to get cited by AI answer engines?
Five things, in order: crawler access, index discovery, a liftable passage, a named source behind every number, and repeated cross-engine measurement.
Most published checklists for getting cited by AI stop at naming those steps. The column that matters is the third one below, because a step nobody can verify is a step nobody can be sure is finished. Every row here ends in an observation someone can make today rather than a state of mind.
| Step | What to do | The check that proves it |
|---|---|---|
| 1. Crawler access | Allow GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, PerplexityBot, Googlebot and Bingbot in robots.txt, and serve the content inside the HTML the server returns | Request the live URL once per user agent and confirm both a 200 response and the body text inside it |
| 2. Discovery | Submit the URL through Search Console and IndexNow, keep it in a sitemap with an honest lastmod, and earn references from pages that are already indexed | Ask each engine for the page by name and by URL - a fetch on demand is not the same as a search result |
| 3. A liftable passage | One self-contained answer under the question heading with no links inside it, plus structured data - Article, FAQPage, BreadcrumbList - that matches the visible text | Read the passage on its own: if it needs the sentence above it to make sense, an engine cannot lift it |
| 4. A named source | Every figure carries the source, the date and the population it was measured on | Trace each number back to the primary study and confirm what was actually counted |
| 5. Measurement | A fixed prompt set run across ChatGPT, Claude, Perplexity and Google AI Overviews on a schedule | A dated table of mentions and cited URLs that can be compared with the previous run |
Which crawlers do I need to allow?
Seven user agents matter, and blocking any one of the seven hides the page from that engine no matter how good the content is.
The seven are GPTBot and OAI-SearchBot from OpenAI, ClaudeBot and Claude-SearchBot from Anthropic, PerplexityBot, Googlebot and Bingbot. Two conditions travel with them. The page has to be server-rendered, with the answer present in the HTML the server returns, because a passage assembled by JavaScript after load is not reliably part of what a retrieval crawler stores. And a CDN or firewall rule can deny a bot that robots.txt allows, which is why the check is a real request rather than a reading of the config. The crawler names and the retrieval routes differ per engine: see what decides a ChatGPT citation, how freshness works on Perplexity, the Anthropic crawler split, and what Google says to ignore.
One item that sits on nearly every crawler-access checklist is worth naming plainly, because it is widely sold as step one. Google's own documentation states that Search does not use llms.txt or similar files, and that keeping one neither helps nor harms visibility there. Running one is defensible for agentic clients that do read it, and this site runs one for that reason. Presenting it as the way engines find your best pages is not supported by anything published.
Why is an accessible page still not cited?
Access and discovery are separate systems. A crawler can fetch a page successfully while no engine index holds the URL, so nothing ever gets quoted.
This site is the worked example. On 30 August 2026 we requested one of our own live URLs with four user agents in turn: a normal browser string, OAI-SearchBot, PerplexityBot and ClaudeBot. All four received a 200 response with the full page body. Eight days earlier, one engine's web search had reported that it could not find the site at all, then described it accurately the moment the URL was handed to it directly. Nothing about access was broken. The domain was simply not present in the index that engine searches.
The practical consequence reorders the usual advice. Crawler access is necessary and cheap, and it is where every checklist starts, but it buys nothing on its own. What moves a URL into an index is submission through the channels an engine actually reads, a sitemap that tells the truth about what changed, and references from pages already sitting in that index. On a young domain that third one does most of the work, and it is the slowest of the three by a wide margin.
Should the answer passage be 40 to 60 words?
Shorter, most likely. The audit behind the 40-to-60 recommendation measured passages of 120 to 150 characters, roughly 20 to 25 words.
The 40-to-60 figure is a genuine market consensus. It appears in the AI Overviews written on this topic and in most competing guides. But the audit that people cite for it measured something narrower. Across fifteen domains, 72.4% of the blog posts ChatGPT cited contained an answer passage. The auditor defined that passage as a self-contained explanation of 120 to 150 characters placed after the title or after a question-based heading. That is a finding about cited blog posts on fifteen domains, not about every page an engine quotes, and it points at a passage roughly half the recommended length.
A second number from the same audit settles a contradiction the market has not noticed it is carrying. Google's AI Overview on this topic, read in August 2026, asks for both at once: keep the answer passage link-free, and ground the page with ten to fifteen outbound links to trusted sources. In the audited set, about 91% of the passages that got cited contained no links at all. Both instructions can be satisfied at once, and the resolution is placement rather than compromise: the passage stays clean, and the links, sources and caveats go in the paragraph directly beneath it. Length, placement and worked examples of answer-first writing are in the guide to writing one. The markup that describes a passage sits in the schema layer, and the way to present a figure inside it is in the note on sourcing every number.
How long until a page gets cited by AI?
No reliable public number exists. Vendor timelines circulate without a method or a sample, and engine indexes refresh on their own schedule.
We used to quote a four-to-eight-week range here, sourced to a vendor FAQ. It came with no sample size, no method and no measurement window, and the same vendor publishes a wider spread elsewhere on its own site. A number like that is a marketing estimate wearing the clothes of a finding, and it does not belong on a page that asks readers to trace every figure to its population.
What can be timed is the sequence rather than the outcome. A crawler request appears in server logs within hours or days of submission. Index membership can be tested directly by asking each engine for the page. Being cited by AI, when it happens, follows index membership rather than publication. It then moves as the underlying index refreshes. That is why the honest answer to the timing question is three dated checks rather than a range of weeks.
How do you measure AI citations?
Run a fixed prompt set across the four answer engines on a schedule, recording mentions and cited URLs as two separate numbers.
The two numbers behave differently and mixing them hides the problem. A brand mention means the engine resolved the entity; a cited URL means it chose the page as a source. A brand can score well on the first and zero on the second, and that gap points at discovery rather than at content. Keep the prompt set fixed between runs, otherwise nothing is comparable. We call the resulting metric Share of Model.
Citations also decay. Engines re-retrieve, competitors publish, and indexes refresh, so a page cited in one run can be absent from the next without anything on it changing. That is an argument for a repeating schedule rather than a launch audit, and for refreshing the pages that lose citations first. For the same techniques sorted by how much published evidence stands behind each one, rather than in execution order, see the strategies ranked by the evidence behind them.
Automating the checklist
Running five steps across one page is an afternoon. Running them across a site, and re-running the checks every time the content or the indexes move, is a standing job. Citematic automates this end to end and reports Share of Model per engine, so the pages losing citations surface on their own. See plans.
Key takeaways
- Crawler access is necessary, cheap and insufficient: a page can return 200 to every bot and still never be cited by AI.
- Discovery is the step most checklists skip: submission, an honest sitemap, and references from already-indexed pages.
- The audit behind the standard advice measured answer passages of 120 to 150 characters, about half the length it recommends.
- Roughly nine in ten cited passages carried no links, so put the links in the paragraph below rather than dropping them.
- No credible public timeline for a first citation exists; measure the sequence with dated checks instead.
Frequently asked questions
Which engines does this checklist cover?
ChatGPT, Claude, Perplexity and Google AI Overviews. The five steps are the same across all four because they describe the same pipeline: reach the page, hold it in an index, lift a passage from it. What differs per engine is the crawler name, the retrieval channel, and how visibly the citation is displayed in the answer.
How do I check that an AI crawler can reach my page?
Request the live URL once per user agent - GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot, PerplexityBot, Googlebot, Bingbot - and check two things in the response: a 200 status code and the actual body text. A 200 with an empty or JavaScript-dependent body means the crawler arrived and got nothing worth quoting.
Where do the links go if the answer passage cannot contain them?
In the paragraph directly beneath it. The passage answers the question and stands alone; the paragraph below it carries the internal links, the source citation and the nuance. This ordering costs nothing in usefulness to a human reader and keeps the quotable unit clean.
Should I build a separate page for each AI engine?
Only when the engines genuinely differ on that topic - crawler behaviour, retrieval channel, citation display. Duplicating one checklist five times with the engine name swapped is thin content, and the engines cite whichever version they already hold. One checklist plus per-engine pages for the real differences is the workable split.
How is getting cited different from SEO?
SEO wins a ranked link; getting cited by AI wins a quoted sentence inside the answer, often with no click attached. The groundwork overlaps almost entirely - crawlable, indexed, well-structured pages - but the target differs, so the unit of success and the way you measure it both change.
Sources
- Adam Gnuse, Search Engine Land, How to get cited by ChatGPT: The content traits LLMs quote most. Source of the 72.4% and the roughly 91% figures used above. Population: blog posts on fifteen audited domains that were indexed in Google and measurable in GA4, not all pages ChatGPT cites. Published November 19, 2025. Checked August 30, 2026.
- Google Search Central, AI features and your website. Source of the statement that Search does not use llms.txt, and that indexing plus snippet eligibility are the stated conditions for appearing in AI features. Checked August 25, 2026.