okaihub

How to Get Your AI Product Cited by ChatGPT and Perplexity

The two stages of an AI citation — retrieval then synthesis — and the concrete changes that affect each, in order of how much they matter.

Published 2026-08-14 · 5 min read

Being cited by an AI assistant is two separate problems, and most advice about it conflates them. First the system has to find your page. Then, given several candidate pages, it has to choose to quote yours. Fixing the second without the first does nothing.

Here is both stages, ordered by how much each actually moves the outcome.

Stage one: can it find you at all

This stage is binary and almost entirely under your control. Most sites fail it.

  1. 1

    Permit the retrieval crawlers in robots.txt

    Explicitly allow the bots that fetch at answer time, separately from those that train on data. Most sites block or ignore both. Allowing them is the single highest-leverage change available and it takes ten minutes.

  2. 2

    Make sure content is server-rendered

    Text that appears only after JavaScript executes may never be seen. Render the substance on the server and enhance it afterwards if you must.

  3. 3

    Remove access barriers from important pages

    A login wall, an aggressive cookie banner, or a bot challenge on your product pages will block retrieval entirely. Keep those for the app, not for the pages that describe it.

  4. 4

    Get into the sitemap and submit it

    Not sufficient on its own, but its absence on a small domain means pages are found only by following links.

  5. 5

    Publish a machine-readable summary

    An llms.txt-style file describing the site and linking its important pages gives a system a small structured document to retrieve instead of crawling. Adoption is still low enough that having one is an advantage.

Stage two: being chosen over the alternatives

Once retrieval has several candidate pages, the model synthesises. What makes a page quotable:

PropertyWhy it mattersHow to check yours
Definition-first openingThe first sentence gets lifted verbatim. If it needs context to make sense, it will not be usedRead your first sentence alone, out of context
Self-contained factsA number or claim with its scope attached can be quoted safelyWould the sentence be wrong if quoted on its own?
Real table markupStructured data in tables extracts reliably; the same data in styled divs often does notView source and look for table elements
A visible, accurate dateFreshness is weighed explicitly in retrieval and synthesisIs the date real? Inflated freshness destroys trust when noticed
Consistent entity namingName, domain and profiles identical everywhere, so mentions resolve to one productCompare your site, listings and social profiles
Specificity over comprehensivenessA page that answers one question precisely beats one that covers five vaguelyCould this page be split? Then split it

Stage two-and-a-half: being in the sources it reads

Models do not only retrieve your site. They retrieve whatever pages discuss your category — directories, comparison articles, roundups, forum threads. Being named accurately on those pages matters as much as your own site, and for a new domain it often matters more, because those pages already have authority.

  • Directory listings. Consistent name, description and links across every listing. This is how a system learns your product exists in a category, and inconsistency across listings is the main reason attribution fails.
  • Comparison and alternative pages. Being included in an honest comparison — including ones you do not control — is a strong retrieval signal. Where one is inaccurate about you, ask for a correction; most authors will make it.
  • Community threads. Technical discussions are retrieved for "how do I" questions. A substantive answer you wrote years ago can surface.
  • Your own comparison pages. If you write the honest comparison against the incumbent, you become retrievable for that query rather than leaving it to whoever writes it first.

The order to do this in

  1. 1

    Week one: permissions and rendering

    robots.txt, server rendering, sitemap, and removing access barriers from public pages. Binary fixes with the largest effect.

  2. 2

    Week two: the definition pages

    Rewrite your product page and your most important category page so the first sentence defines the thing. This is where quotation happens.

  3. 3

    Week three: structure and entities

    Convert key facts into real tables. Add sameAs links to your profiles. Verify your name, domain and description match across every listing.

  4. 4

    Week four: machine-readable summary

    Publish an llms.txt-style file listing your important pages. Link it from the footer so it is discoverable by humans too.

  5. 5

    Ongoing: presence in retrieved sources

    Directory listings maintained, comparison pages written, community questions answered substantively.

How to check whether any of it worked

There is no dashboard. The method is manual and should be scheduled rather than done once:

  • Pick eight to twelve questions a buyer would ask — "best tool for X", "X alternatives", "how do I do X", "what is X".
  • Ask each assistant the same set monthly and record whether you appear, and how you are described. The description matters as much as the appearance; being cited with the wrong framing is worse than not being cited.
  • Segment analytics by referrer for the assistant domains. Most strip UTM parameters, so referrer is the only reliable signal, and it will undercount because some traffic arrives with no referrer.
  • Watch crawler hits in server logs, by user agent. Crude, but it shows whether retrieval is reaching you at all — and if it is not, no amount of writing improvement will help.

What does not work

  • Keyword density. A model synthesising an answer is not matching terms the way an index does.
  • Hidden instructions in page text. Attempts to direct a model via invisible text are actively filtered, and the same infrastructure that catches them also handles search penalties.
  • Paying for inclusion. There is no general mechanism for buying a citation. Anything sold as one is selling something else.
  • Volume of pages. Ten thin pages do not beat one precise one, and publishing at scale is the pattern search policies penalise.
  • Waiting for authority. Inclusion depends less on domain authority than a search ranking does. A precise, reachable page can be cited ahead of a larger competitor with a vaguer one — which is the main reason this channel is worth a new product's attention.

Retrieval is a permissions problem and synthesis is a clarity problem. Most sites have both wrong and fix neither, because the fixes are unglamorous: a robots file and a first sentence.

More in SEO & GEO