How to get your pages cited by ChatGPT: a page-level checklist

How ChatGPT chooses pages to cite

When search is on, ChatGPT turns the question into one or more web searches, fetches a handful of candidate pages, and writes an answer with links to the ones it used. The candidates come from a search index, so the page has to rank somewhere for the queries ChatGPT runs. The citation goes to the page whose text most directly supports a sentence in the answer. How ChatGPT decides which pages to cite goes into the mechanism; this article is the practical consequence.

Two things follow. First, a page nobody can find in search will not be cited, however good it is. Second, among the pages it does find, ChatGPT favours the ones it can quote cleanly: a clear claim, in plain text, with the context in the same paragraph. The checklist below is built around those two facts.

The page-level checklist

Work through this for one page at a time, starting with the page you most want cited for a question you have already chosen. Tick each item; the "why" tells you what you are buying with it.

1. OAI-SearchBot is allowed in robots.txt. OpenAI's documentation lists OAI-SearchBot as the crawler for ChatGPT search, separate from GPTBot (used for training) and ChatGPT-User (fetches when a user asks). Many sites blocked GPTBot in 2023 and copied the rule to every OpenAI agent; if OAI-SearchBot is disallowed, the page cannot be in the index that ChatGPT searches. Check the file, and read AI crawlers explained for the full list of agents.
2. The page returns real HTML to a bot, with no challenge. A CDN's bot protection, a "verify you are human" interstitial, a login wall or a 403 for unknown user agents all make the page invisible. Fetch it with curl and a bot user agent, or read a crawler log, and confirm you get a 200 and the article text.
3. The text is in the HTML, not rendered only by JavaScript. Crawlers that fetch to cite generally read the HTML they receive; a page whose body arrives through a script after load may be read as empty. Server-side rendering or static generation fixes this. View the page source (not the inspector) and check the paragraphs are there.
4. The title states one question and the first paragraph answers it. A page that tries to be about everything is cited for nothing. "How much does a scheduling tool cost per user?" followed by a two-sentence answer with a range and a date is a page an assistant can use. A title like "Pricing insights" is not.
5. The page contains facts that can be quoted. Numbers with their units and dates ("from 8 euros per user per month, as of September 2026"), named features, named integrations, a price where you can show one, a step count, a limit. Assistants attribute sentences to sources; a page of adjectives gives them nothing to attribute.
6. A visible date and a named author. Published and updated dates in the text (not only in metadata) and an author with a role. Freshness matters for time-sensitive questions, and a named author is one of the trust signals search engines' own guidance mentions, which assistants that search inherit.
7. Headings mirror the sub-questions. If a buyer asking the main question would also ask "does it work with Google Calendar" and "is there a free plan", make those the H2 headings, worded as they would ask them. Each section then answers its heading in the first sentence, so a section can be quoted on its own.
8. A short FAQ at the end. Three to five questions, phrased as people type them, each with a two- or three-sentence answer. It gives the page extra quotable passages for the adjacent questions, and the structure is easy to extract.
9. Structured data matches what is on the page. Article with author and dates, FAQPage for the FAQ you actually show, Organization on the site. It does not earn citations by itself (see what schema.org actually does), but it removes ambiguity about what the page is and who wrote it. Never mark up content that is not visible.
10. One canonical URL, no duplicates. The same text at /pricing, /pricing/ and /pricing?ref=nav, or a PDF copy, or a syndicated version, splits the signals and lets the assistant cite the wrong one. Set rel="canonical", redirect the variants, and keep one version.
11. Fast, simple HTML. A page that takes seconds to respond, or carries megabytes of scripts, is more likely to be skipped or truncated by a crawler on a time budget. Keep the main content near the top of the HTML and the page light.
12. The page is linked from pages that already get cited. Look at which of your pages ChatGPT cites today (ask it, or check your citations report) and link the new page from them, in the body, with descriptive anchor text. Then do the same off-site: a mention in a comparison article, a listing on a review site, a forum answer that links to it. Those pages are already in the candidate set; a link from them puts yours next to them.

The last item is the one most teams skip and the one that most often moves the result. A perfect page with no links is a page the search step never surfaces.

What the checklist leaves out

Word count targets, keyword density and "AI-friendly" phrasing. None of those is documented by anyone to matter, and chasing them produces worse pages. Length should be whatever answering the question takes; a 400-word page that nails one question beats a 3,000-word one that circles it.

The site level

One page can be cited for one question. A site gets cited across a topic when it covers the topic. Three things to add once the page-level work is done.

Topic coverage

List the 20 to 50 questions your buyers ask on the way to a purchase (how to choose them), and check which have a page that answers them. The gaps are the content plan. A site that answers the problem, the category, the comparison, the price and the how-to questions gives an assistant a reason to come back for each.

Comparison pages

"X vs Y" and "best X for Z" are the questions where assistants most often cite a vendor page, because buyers ask them and third-party pages are often thin. Write honest ones: a table of the differences, where the other product is better, who should pick which. A comparison that only praises you reads as an ad and gets cited less. How AI uses best-of lists and comparison pages has the detail.

A glossary

Short definition pages for the terms in your category ("what is round-robin scheduling") are cheap, get cited for definition questions and give the rest of the site something to link to. Each one is one question, one answer, 150 to 300 words, dated.

How to test whether it worked

Testing is manual, and it only means something when repeated.

1. Ask the real question in ChatGPT with search on. Use the exact phrasing a buyer would, not your keyword. Use a temporary chat so your history does not colour the answer.
2. Note the citations. Which pages, from which sites, in what order, and which sentence each supports. Write them down with the date.
3. Repeat across days. Answers vary between runs and change as the index updates. Ask the same question on five different days before you conclude anything.
4. Vary the phrasing slightly. "best scheduling tool for agencies" and "what scheduling tool should an agency use" can pull different sources. Track both.
5. Compare with the other assistants. Perplexity, Gemini and Google's AI Overviews search differently and cite differently; a page that works in one may not in another. Getting cited by Perplexity covers the differences.

What to look for over four to six weeks: your page appearing in the citations for its question at all, then more often than not, then higher in the list. If after six weeks it is never cited while the same third-party pages always are, the page is not the problem; the lack of links from those pages is. Go back to item 12.

If you are doing this by hand for more than a handful of questions, keep a spreadsheet with one row per question per day. Or use a tool that runs the prompts on a schedule and records the citations; that is what Echo's Citations page does.

A worked example

Fernwick, a made-up scheduling tool, wants its page cited for "how much does a scheduling tool cost for a small sales team". The page they had was titled "Pricing insights for modern sales teams", opened with three paragraphs about productivity, and showed prices only in a JavaScript pricing table.

The fixes, in order of the checklist: robots.txt already allowed OAI-SearchBot, but the CDN challenged unknown bots, so they added an exception. They retitled the page "How much does a scheduling tool cost for a small sales team?", wrote a two-sentence answer with a per-user range and "as of September 2026" at the top, and rendered the pricing table as plain HTML. They added H2 headings for "What do the paid plans add?" and "Is there a free plan?", a five-question FAQ with matching FAQPage markup, a byline with the head of marketing's name and a visible updated date. Then they linked the page from their comparison page, which ChatGPT already cited, and asked two review-site listings to point to it.

They tested daily for a month. The page showed up in the citations in week three, after the review-site links went live. The fix that moved it was the last one.