How to Get Cited in ChatGPT and Perplexity: Win the Live Fetch

One dated guide on our site fed 273 verified live answer-time fetches in 30 days; our category and hub pages fed zero. If you want ChatGPT and Perplexity to cite your business, the mistake is treating citations mainly as a training-data problem when a buyer may be getting an answer from live retrieval and clickable sources right now.

I used to tell clients to think in months and hope the next model remembered them. I was wrong. The more useful question is whether an assistant can fetch a page today, find a passage that answers the buyer, and cite it. Our 30-day edge-log breakdown does not prove that every fetch became a citation. It does show where the live reads went. This guide is about building the kind of page those fetchers can use, then checking whether it actually gets quoted.

How does an assistant find you when a buyer asks?

Training data is slow, opaque, and mostly out of your hands. Live retrieval gives you something to work on today. With browsing, ChatGPT can retrieve pages and return 3 to 6 clickable citations. Without browsing, a plausible-looking reference is not the same thing as a live, clickable source. Perplexity makes the retrieval process more explicit: it interprets the query, pulls candidates from its index, reranks them, and synthesizes an answer with inline citations. Its reported pipeline draws 10 to 30 candidates from an index of around 200 billion URLs, then produces 5 to 10 citations.

The buyer’s wording is not the whole target, either. ChatGPT can rewrite a prompt into multiple backend sub-queries. A request for the best tools in 2026 might trigger separate comparison, review, and pricing searches. Your page does not have to repeat the prompt word for word. It has to offer a passage that supports a claim the assistant is trying to answer.

That changes the job. Instead of waiting for a model to remember your brand, make a specific answer available at the moment it looks. Then give it a reason to choose that answer over the other pages it finds. Start with the fetch; clever wording cannot rescue a page the fetcher cannot read.

What did our 30 days of edge logs actually show?

The source is our own edge logs, not a visibility-dashboard estimate. We counted claimed AI bot hits over 30 days, then kept the fetches that behaved like live answer-time reads. Claimed crawler traffic often failed that check: the wrong agent for the job, no follow-through fetch, or archive-style crawling rather than a read associated with an answer. The full breakdown explains the method.

The limit is just as important as the gap in the table. This is one site over one month, not a citation-rate benchmark or proof that dating a page alone causes retrieval. The logs count fetches, not buyers or revenue. What they give me is a useful priority: on the same domain, narrow pages with current, extractable answers drew live reads while broad category pages did not. I would fix the latter’s answer before writing another broad page.

Step 1: Can the fetcher read the answer without your browser?

OpenAI sends different agents for different jobs. Do not confuse training access with search access. GPTBot collects training data; OAI-SearchBot supports the ChatGPT search index; ChatGPT-User fetches a page on demand. Blocking GPTBot in robots.txt does not by itself remove you from search answers, but blocking OAI-SearchBot can cut off the index those answers use. The crawler roles matter. I have seen sites block everything with one wildcard disallow, then spend months improving content no assistant could retrieve. Check the file before rewriting a page.

For Perplexity, the distinction is similar but the access checks do not end at robots.txt. PerplexityBot respects that file and is documented as separate from foundation-model training; Perplexity-User fetches on a user’s request. A firewall challenge or answer text available only after JavaScript runs can still get between the fetcher and the passage, even when the robots file looks fine. That is why server-side HTML and firewall rules belong in the same check.

Run the check in this order:

1. Read robots.txt. Confirm the agent you need is not blocked by a broad rule.
2. View the page source. Find the answer itself in the raw HTML, not just in the rendered page you see after scripts run.
3. Try the page without scripts. Check whether the useful text remains available.
4. Check the firewall. Make sure verified fetchers receive the page rather than a challenge.

The 45-minute fetch check is a useful companion if you want to work through those points before touching copy. A fetchable page is not guaranteed a citation. An unreadable one gives the assistant little to cite.

Step 2: Which buyer questions deserve a page?

Do not begin with 50 prompts and a spreadsheet that looks impressive until someone asks what the prompts mean. Start with 10 to 15 questions buyers ask close to a purchase. Listen for cost, timing, fit, and alternatives: what does it cost, how long does it take, which option fits a business my size, who does this well in my city? Pull candidates from closed-won calls and the search terms report before opening a keyword tool.

Say you spend $20k a month on ads. “What is PPC?” is not the question I would build around. “How much does Google Ads management cost for ecommerce at $20k spend?” and “Who manages Shopping feeds for margin-based bidding?” are closer to a decision. Each asks for something a page can answer directly rather than a paragraph of category copy. Because an assistant may split a buyer’s prompt into sub-queries, a passage with a real pricing answer can be more useful than a page that repeats the buyer’s wording five times.

Write the questions as buyers say them. Then cut the ones you cannot yet answer with evidence. If you have no date, price range, process step, or result to put on the page, skip that question for now. Do not disguise an empty answer with a longer introduction.

Keep the surviving prompt set stable when you measure it. A team tracking AI visibility reported moving from 50 citations in June 2026 to 121 in August 2026 as the same pages were cited more often, rather than by publishing more pages. That is their result, not a target I would promise you. The useful discipline is keeping the questions consistent enough to see whether your pages start appearing.

Ten questions you can prove beat forty you cannot.

Step 3: What belongs in the passage an assistant might quote?

Put the direct answer near the top, within the first 150 words. Date it when the answer depends on current conditions, and give the reader the number or condition they came for. One analysis reports that 44% of citations come from the first third of a page. That does not make the rest disposable. It does make a slow introduction expensive.

I build around one self-contained answer block of roughly 200 words. It states the answer, then the price range or timeline if there is one, then two or three conditions, then who the answer fits. A clear question heading, a table or list where it helps, and FAQ schema can make the page easier to parse. The point is not to decorate a thin claim with formatting. It is to let someone lift a complete passage without hunting through five sections to understand its limits.

Put product copy and cross-sells below that block. Our category pages had plenty of text, but no single question answered cleanly. More copy would not have fixed that. A visible date and a changed fact are more useful than a full rewrite that leaves the answer vague; the same citation analysis reports that pages updated within 30 days appear in ChatGPT answers about 3.2 times more often. Treat that as a reason to keep facts current, not to change a timestamp for show.

If you also target Google, the passage-first habit carries over. The companion guide to AI Overviews covers where that surface differs. For this page, stay disciplined: one buyer question, one answer that can stand alone, proof close enough that the claim survives being quoted.

Step 4: Who else can support what your page says?

Your own page can make a claim available. A third party can give the assistant corroboration. In a large Perplexity query study, Reddit accounted for 21.6% of citations and YouTube for 20.8%. Those figures describe that study’s queries, not your category. The same source reports a review pattern on the ChatGPT side: businesses with a real Trustpilot presence averaged 4.6 to 6.3 citations against 1.8 without one. Neither comparison means opening an account produces citations on command.

What I take from them is practical. A detailed outside mention gives a retrieval system something beyond your self-description to weigh. Five directory links with interchangeable descriptions are not the same as one review that says what the buyer hired you to do, when, and with what outcome. Do not buy a pile of empty mentions and call it proof.

Ask five recent buyers for detailed reviews that name the job, timeline, and outcome. Pitch a comparison in your category with real prices and dates rather than feature dots. If a local listicle author covers businesses like yours, give them a number they can print. You cannot write their verdict for them, and you should not try. You can make the underlying claim specific enough to check.

If nobody else can state your claim plainly, give them better proof before asking for another citation.

Step 5: Did the fetch become a citation?

Freeze the 10 to 15 buyer questions from step two. Once a month, run each in a fresh ChatGPT thread and a fresh Perplexity thread. A continuing chat remembers context; it can make an apparent improvement less useful as a comparison. Keep the record simple:

Do not count a brand mention as a page citation. I count a win when the same page is cited for the same question in two consecutive runs. One appearance might be luck; a repeat gives you a better reason to think retrieval is working. It still is not a promise about the next buyer’s answer.

If nothing moves after 30 days, resist the urge to publish more pages. Recheck in order:

1. Fetch: Is OAI-SearchBot allowed? Does the answer still appear in server-side HTML? Is a firewall challenge in the way?
2. Passage: Could a stranger quote the opening answer without reading the rest of the page?
3. Proof: Does any third party state the same claim with useful numbers or detail?

That order matters. A review will not repair an unreadable page, and a clean fetch will not turn vague copy into an answer. Most misses I see are at the first two checks. Let the log tell you which one to fix rather than treating another page as the default solution.

Who should skip this, and what should you do today?

Skip a page you cannot substantiate. If you launched last month, have zero reviews, and no third party has named your price or process, work on fulfillment and ask for five detailed reviews first. Skip the ten-page plan if your category gets ten buyer questions a year, too. One quotable page may still help; ten pages for ten searches is a bad trade.

For everyone else, set expectations by the work, not by a model-training cycle. A fetch fix and an answer-first rewrite can show up as a first citation in two to three weeks because a later live fetch can read the change. Third-party mentions can take 60 to 90 days because someone else has to publish. Neither window guarantees a result. Anyone promising citations in 48 hours is selling you a dashboard.

Do one thing today: pick the buyer question that already makes you money. Check that its page serves the answer in raw HTML and allows OAI-SearchBot, then rewrite the first 150 words as a dated answer with one real number. Run that exact question in fresh ChatGPT and Perplexity threads next week and record who gets cited. Win the fetch first. Then give the assistant a reason to pick you.