How to get your page cited by ChatGPT and Perplexity

More and more buying research now starts in an AI assistant instead of a search box, and the game there is different. An answer engine does not hand back ten blue links; it retrieves a handful of passages, synthesises them into prose, and sometimes credits the sources it used. To be one of those sources your page has to be easy for a machine to find, parse, trust and quote. That is a different job from ranking, and most pages are not set up for it.

What the engine actually does

When someone asks a question, the assistant pulls a few relevant passages from across the web, stitches them into one answer, and may cite the pages it leaned on. So you are not competing for a ranked position, you are competing to be a cleanly liftable passage on the exact thing being asked. Two things follow: write in self-contained chunks, and make the facts unambiguous to a machine.

Let the right crawler in first

None of the signals below matter if the engine cannot fetch the page, and this is where a lot of sites quietly lose. The crawler that reads a page for a live answer is not the same one that reads it to train a model, and the two are controlled separately in robots.txt. OpenAI runs GPTBot to collect training data and OAI-SearchBot to surface and cite pages in ChatGPT search, and its own documentation states the settings are independent, so a site can allow OAI-SearchBot to appear in search results while disallowing GPTBot. The upshot: the common move of blocking GPTBot to keep your content out of training does nothing to your citation odds, but a blanket AI-bot block that also catches the search crawler quietly removes you from the answers.

Perplexity splits the same way. PerplexityBot indexes pages for its answer engine and respects robots.txt, while Perplexity-User fetches a page only when a person asks about it. If you want to be cited, the safe default is to allow the retrieval crawlers, OAI-SearchBot and PerplexityBot, even when you disallow the training ones. So before tuning any of the signals below, read your own robots.txt for a wildcard disallow or a plugin rule that blocks AI user-agents wholesale, because it can shut the front door on the exact engines you are trying to appear in.

robots.txt is not the only gate: check your CDN

There is a second gate above robots.txt, and it is the one people miss. robots.txt is a rule the crawler reads only after its request reaches your site, so it decides nothing if the request is stopped before that. A lot of sites sit behind a CDN or firewall that can block a bot by user-agent at the network edge, before robots.txt is ever fetched, and that edge rule overrides whatever your robots.txt allows. Cloudflare, which sits in front of a large share of the web, made this concrete in July 2026: it sorts AI crawlers into three categories, Training, Agent and Search, and from 15 September 2026, on pages that display ads, it blocks Training and Agent by default while leaving Search allowed by default. That new default applies to sites onboarding after that date, and existing customers can set their own per-category preference first.

The good news for citations is that the Search category, the one that surfaces and cites you in an answer, stays allowed by that default. The risk is the broader block: a managed rule or a block-all-AI-bots toggle you switched on yourself can catch the Search crawler too, and because it runs at the edge a perfect robots.txt allow will not save it. So when you audit access, check the layer above robots.txt as well, and confirm your CDN or firewall is not stopping OAI-SearchBot or PerplexityBot before the request reaches your origin. If you are not being cited despite an open robots.txt, this is the first place to look.

Say it once: the content-signals line

Managing each crawler by user-agent works, but it means keeping a list of bot names current as new ones appear, and allowing or blocking a fetch says nothing about what a crawler may do with the page once it has it. There is now a tidier way to state the intent in one place. In September 2025 Cloudflare published a Content Signals Policy, an addition to robots.txt that lets you express what a crawler may do with a page after it fetches it, not just whether it may fetch it. It defines three signals: search, for building a search index and returning links and short excerpts; ai-input, for feeding your content into an AI model to answer a question in real time, the retrieval that grounds an answer engine; and ai-train, for training or fine-tuning a model. You set each to yes or no, or leave it unset to say nothing, on a single line such as Content-Signal: search=yes, ai-train=no. That line says be found and cited, do not train on me, which is exactly the split most publishers want.

Two honest caveats keep this from being a magic switch. First, it is a stated preference, not a lock: like robots.txt itself it works only for crawlers that choose to honour it, so it does not replace the edge rules or the user-agent controls above, it sits alongside them (Cloudflare notes the signals may carry legal weight in some jurisdictions, but technically nothing stops a crawler that ignores them). Second, watch the default if you sit behind Cloudflare: for managed robots.txt customers it now writes Content-Signal: search=yes, ai-train=no for you and leaves ai-input unset. That default keeps you citable, but read your own robots.txt to confirm it matches what you want, especially the ai-input line, which is the one that decides whether an answer engine may pull your page in to ground its reply.

The signals that get you quoted

• Descriptive headings: a subheading that names the question it answers splits your page into passages that can each stand alone as an answer, which is exactly what gets lifted.
• FAQ blocks: real questions with short, direct answers map almost one-to-one onto the way people prompt an assistant.
• Citable stats: a specific, attributable figure is the kind of thing an assistant likes to quote, because it is concrete and checkable.
• Structured data (JSON-LD): schema removes ambiguity about what your page describes, so the model is more confident you are relevant.
• Freshness: a visible update date and the current year tell a model the page is not stale, which it weighs when choosing what to cite.
• An llms.txt file: a young convention, but a cheap one, a curated map of your best pages for the crawlers that read it.

Entity clarity comes first

Before any of the formatting, the model has to be sure what and who your page is about. Name the product, company or topic plainly and consistently instead of leaning on pronouns and clever phrasing. Entity clarity, schema and citable stats are usually the three that move the needle most, so fix those before the rest.

Past citation: being buyable by the agent

The endpoint of research that starts in an assistant is the purchase, and a growing slice of it now completes inside the assistant instead of on your site. ChatGPT added Instant Checkout, where a shopper buys without leaving the chat, powered by the Agentic Commerce Protocol, an open standard OpenAI built with Stripe and published for any merchant to use with its existing payment provider. It launched with US Etsy sellers and is rolling out to over one million Shopify merchants. Being buyable is a separate capability from being citable: you can be quoted in an answer with no way for the agent to close the sale, or be transactable and rarely cited. Citation gets you into the answer; being on a supported commerce surface gets you the buy button.

Two things make this worth preparing for rather than dismissing. First, it is becoming infrastructure, not one vendor experiment: at the Visa Payments Forum on 10 June 2026 the card networks set out their own rails for agent-initiated payments, Visa detailing tokenised payments for agent checkout with OpenAI under a Trusted Agent Protocol aligned to the same open standard, and Mastercard its Agent Pay rails alongside. Second, the reassuring part: the merchant stays the merchant of record. You keep your own checkout terms, pricing, fulfilment, tax and chargebacks and the customer relationship, so the assistant is a new front door rather than a new owner, and OpenAI says results are not influenced by whether checkout is turned on and buyers do not pay more, so the citation signals above still decide whether the agent surfaces you first. The honest caveat is that this is early and moved fast through 2026, so treat being buyable as an emerging channel to get ready for, and keep the discovery fundamentals, because the agent has to find and trust you before it can buy from you.

Measure it, do not guess

None of this guarantees a spot. The engines change how they retrieve and cite often, results vary from one run to the next and by region, and the honest move is to treat the signals as a checklist of things that genuinely help, then track whether they work. Run a fixed set of buyer prompts each month and record, for each, whether you were mentioned, whether you were cited with a link, and roughly where in the answer. That grid is your AI-visibility trend line. The tool generates those prompts, scores your page on every signal above, and hands you the fixes in priority order.