How to get cited by ChatGPT and Perplexity

How ChatGPT and Perplexity choose their sources

Citations come from retrieval. At answer time the engine searches an index of the web, fetches a few pages live, and links what it used. Two conditions have to hold for your page to appear: the engine's crawler can reach it, and it contains a passage that answers the question.

Half of getting cited is plumbing, and it is the half teams skip. Every engine publishes which bots it runs and what each one does, and the bots are separate controls: allowing a search crawler says nothing about allowing your content into training. The table below is the access list worth checking on your own domain this week.

Access is necessary and not sufficient. The engine still has to find a passage that completes the prompt, which is where the writing starts.

The first 40 to 60 words decide what gets quoted

Retrieval works on chunks: the engine holds a question and looks for a block that answers it. Put the direct answer in the first 40 to 60 words of the section that addresses the prompt, with the supporting detail after.

Pages lose citations in their opening lines. A section that starts with the company history, a story about a customer, or a hedge such as “it depends on your stack” buries the claim three hundred words down, and the extractor never reaches it. The reliable shape is a heading phrased as the question, one short paragraph that answers it, then the nuance.

Three rules make the block liftable: one claim per paragraph, no pronouns that point back to the previous paragraph, and no qualifiers that only make sense with surrounding context.

What makes a claim worth quoting

A passage gets quoted when it hands the answer a fact it can restate: a number, a definition, a price, a limit, a date. Vague praise gives the engine nothing to carry over.

Compare “affordable pricing for teams of every size” with “$12 per user per month”. Compare “fast support” with “support replies within one business day”. The first version in each pair describes a feeling; the second can be checked, and checkable claims survive being lifted out of context.

Two more shapes earn citations. A one-sentence definition of your category on the homepage (“Northwind is a managed pipeline for moving event data into a warehouse”) answers the definitional prompts that engines answer often. And a scope statement about what you do not do keeps your claims consistent with what third parties write about you, which matters because engines cross-check sources and a page that disagrees with the rest of the web loses.

Consistency is the unglamorous part: the same price on the pricing page, in the docs, in the help center and on your directory profiles. When two of your own pages disagree, the engine picks one and you do not control which.

Your own page is not enough

Most answers to recommendation prompts are built from pages you do not own. Review sites, forum threads, documentation and news carry as much weight as your site, often more, because a brand describing itself does not answer “which one should I buy”.

An analysis of 30 million sources reported by Search Engine Land found Reddit the most cited domain across ChatGPT, Google AI Mode, Gemini, Perplexity and AI Overviews, with YouTube, LinkedIn, Wikipedia and Forbes completing the top five. Review platforms such as Yelp and G2 appeared often in recommendation queries. None of those domains are yours, and that is the point: engines want a second opinion.

So a share of the work happens away from your site. Claim the listing on the review platforms your buyers read and keep its facts current. Write documentation that a third party can link when it explains your product. Get named in the comparisons and listicles your category already reads, either by pitching the author an update or by publishing the specs they copy. Where forum threads complain about you, reply if the rules allow it with the current fact, because that reply can become the source the next answer cites.

This advice is weakest in two situations. On definitional prompts (“what is X”, “how does X work”) first-party docs and encyclopedic pages carry the answer, so your own writing is the lever. And in brand-new categories with almost no coverage outside your domain, there are no third-party sources to earn yet, which means your documentation has to do the whole job. In established categories, expect both fronts. The category map is in which sources AI engines cite.

Freshness: update the fact, not the date

Engines re-retrieve on every question, and the pages that win tend to carry current, checkable information. Update a page when the facts on it change, and stamp it with a date that is true.

The refresh list is short: prices, plan names, limits, version numbers, integrations, headcount claims and anything a changelog already knows about. An old pricing page is the most common stale source on a company's own site, and it keeps feeding answers long after the numbers moved. The same applies to a 2022 blog post that states a feature you retired.

Do not fake freshness. A script that rewrites the publish date while the content stays wrong is worse than an accurate old date, because readers and engines both compare the stamp against what the page actually says. Change the content, then let the date tell the truth about when you did it.

A 30-day procedure worth running

One month is enough to fix access, rewrite the pages most likely to be quoted, and start a measurement loop. It is not enough to move mention counts much, which take longer than people expect.

1. Days 1 to 2: prompts and runs. Write 15 to 25 questions your buyers actually ask (category, comparison, pricing, and a few about your brand), then start daily runs so every prompt has a sample behind it.
2. Days 1 to 3: crawl access. Check robots.txt for OAI-SearchBot, PerplexityBot and Claude-SearchBot, check that no firewall or WAF rule blocks their published IP ranges, and confirm your pages return their text in the HTML rather than after a script runs.
3. Days 4 to 7: answer-first rewrites. Pick the five pages you most want quoted for those prompts (docs, pricing, a category explainer) and move the direct answer to the top of each matching section.
4. Days 8 to 14: make claims specific. Replace praise with numbers, add the definition sentence, add scope statements, and align the same facts across pricing, docs and your profiles.
5. Days 15 to 21: source work. Look at which domains appear in your category's answers, then claim the listings, correct the stale profiles and pitch the comparisons you can reach.
6. Days 22 to 29: publish one citable thing. A benchmark, a dataset, a pricing table other writers can copy. Specific numbers travel further than opinions.
7. Day 30: read the data. Visibility per prompt, which of your pages appear in the sources, and which prompts still hand the answer to a competitor. The full version of this month is in the 30-day GEO starter plan.

How to measure whether citations are moving

Track two things: how often answers mention your brand (Visibility) and which domains those answers link. Everything else is detail until those two are stable.

In advisible the first is on the overview, always with its sample size underneath (“based on 240 answers”), and the second is the Sources view: the domains feeding your answers with counts, exportable as CSV so you can hand the list to whoever does outreach. Daily runs on ChatGPT, Perplexity, Gemini, Grok and Claude keep the sample growing, the answer archive keeps the original text for reading the exact wording, and Run now gives you a fresh read minutes after you publish instead of waiting for tomorrow's schedule.

Read it weekly. A single answer is an anecdote because the same prompt returns different answers on different days, and a week of runs is the first point where a change is worth arguing about. What the numbers mean is covered in reading citation and source data.

What does not work

Keyword stuffing aimed at engines, blocking the search bots and buying placements on cheap listicles all fail for the same reason: retrieval reads what the page says, not what you intended.

• Stuffing the target phrase into meta tags and hidden text. Engines extract from readable text that completes the prompt. A stuffed page gets retrieved and skipped.
• Blocking OAI-SearchBot or PerplexityBot. OpenAI's docs state that sites opting out of OAI-SearchBot are not shown in ChatGPT search answers. You remove yourself from the surface you were trying to win.
• Publishing many near-identical pages. Duplicates split retrieval and add no new claim to quote.
• Q&A blocks with nothing in them. Schema markup and FAQ sections are fine as structure; an unsupported question mark heading with a padded answer gives the engine nothing to repeat.
• Treating one good screenshot as proof. Answers vary run to run, so a day where you appeared in every prompt tells you about that day.