The idea and why it works
When ChatGPT answers a question with web search, it runs extra searches behind the scenes (Ahrefs calls these "fan-out" queries), and those searches increasingly go straight to the source. Publish the original data in your niche (your prices, specs, test results, benchmarks) on stable, undated pages on your own domain, keep them current every quarter, and your page is the one ChatGPT looks up when it checks those facts.
Two studies from the past month point the same way. Ahrefs analysed 66.6 million ChatGPT responses and about 15 million fan-out queries (12 October 2026). It found 6.8% of fan-out queries use the site: operator, and among the roughly 160 most-searched sites, 75.9% of site: searches target brand sites. In plain terms, ChatGPT is going to your domain by name to check facts. Meanwhile aggregator-style sources are fading: Reddit went from about 1 in 5 US answers in March to about 1 in 43 by late September, while PubMed climbed from 64th to 4th among the most-cited domains. In a 20,000-prompt US test, 84.4% of factual answers cited at least one official source.
Greenflag Digital's re-run of its finance study (21 September 2026) shows what happens to pages that coast. Of the 215 pages ChatGPT cited most in October 2025, 107 now get no citations. First-party pages from banks, card issuers and airlines rose from 1.8% to 22.4% of top-500 citations, and pages with a year or date in the URL lost nearly all of theirs, while living reference pages on stable URLs grew.
Two caveats. Ahrefs sells Brand Radar, a citation-tracking tool, and its article promotes it; Greenflag's study also draws on Brand Radar data. And Greenflag covers US personal finance only. Treat both as strong directional signals, not laws.
This works alongside formatting. We covered the markdown trick that helps LLMs parse and cite you, which deals with how your page reads. This play deals with which page gets chosen.
Who it's for
Founders with a product that has prices, plans, specs or limits that people ask about.
Creators and reviewers who run hands-on tests (tools, gear, prompts, templates).
Anyone sitting on first-party numbers: survey results, usage stats, benchmark runs, cost breakdowns.
What you need (tools, time, cost)
A page on your own domain with a permanent, undated URL. Any CMS works.
One dataset you own. A spreadsheet of your prices and specs, or test results you can rerun.
Time: roughly a day to build the first page, then an hour or two each quarter to refresh it.
Cost: nothing beyond your existing site. Paid citation trackers are optional; a manual monthly prompt check works to start.
Step-by-step
List the factual questions people ask about your niche. Price, limits, compatibility, "how fast", "how much", "which is cheaper". Pick the ones only you can answer with authority.
Choose one dataset to own. Your official pricing and specs are the easiest start. A repeatable test (load times, output quality, battery life, cost per task) is the strongest.
Create a stable URL with no date in it. Use /pricing, /specs or /benchmarks/email-tools, never /2026-pricing or /best-tools-may-19.
Put the facts in plain text, near the top. Ahrefs recommends publishing prices, specs and support details in plain text on a page crawlers can find. Use a simple HTML table, not an image or a JavaScript widget.
Show your method and a "last updated" line. The date goes on the page, not in the URL. Explain how you measured, when, and what changed.
Link to it from everywhere. Point your blog posts, docs and newsletter at the living page so it becomes the canonical home of the numbers.
Refresh quarterly and keep the URL. Update the figures, add a changelog entry, never redirect to a new dated version.
Template / example
Here is an outline for a living data page, using a solo SaaS as the example:
URL: yoursite.com/benchmarks/transcription-accuracy
H1: Transcription accuracy benchmark: [Your tool] vs alternatives
Summary (2–3 sentences): what was tested, headline result
Last updated: [date] · Next refresh: [quarter]
Table: tool | plan price | word error rate | minutes tested | test date
Method: audio set, conditions, how errors were counted
What changed since last quarter: bullet list
Raw data: downloadable CSV
FAQ: 4–6 factual questions answered in one or two sentences each
Changelog: dated list of every updateWorked example: a creator selling templates
Say you sell Notion templates and write a newsletter about productivity tools. Readers keep asking which tools have usable free plans. Build yoursite.com/data/productivity-free-plans: one row per tool, with the free-plan limits you checked on each vendor's own pricing page, the date you checked, and a link to that page. Add a column for what you found when you used the free plan yourself, such as whether the export works without upgrading. That column is the part no aggregator has. Under the table, write two or three plain sentences per tool answering the questions readers actually send you. Each quarter, recheck every row, note what changed in the changelog, and mention the update in your newsletter with a link back.
Metrics to watch
Citations in answers. LLM Pulse reported that ChatGPT referral sessions in GA4 fell by around 46% from 9 October, which it attributes to a tracking change that sends visits to Direct. Referral counts are unreliable right now.
Monthly prompt checks. Ask ChatGPT 10–20 of your factual questions and log whether your page is cited.
Brand mentions in answers. Our brand mention monitoring system can be extended to log AI answers.
AI-referred visitors you can still see. A tool like Bear for capturing AI chat traffic helps where referrers survive.
How to check who ChatGPT cites today
Run a baseline before you build anything, then repeat it monthly:
Write 10–20 factual questions a buyer in your niche would ask, phrased the way a person would type them.
Ask each one in a fresh ChatGPT chat with search on, logged out or in a private window so your history does not shape the answer.
Record every cited domain in a spreadsheet: question, date, cited URLs, and whether your site appears.
Mark each citation as official, first-party, publisher or aggregator. The gaps where an aggregator is cited for a fact you own are your first pages to build.
Answers vary between runs, so ask each question twice and count a citation only if it appears consistently.
Pitfalls
Dates in URLs. The clearest loser in Greenflag's data.
Stale numbers. Greenflag advises treating citations as a recurring cost and reviewing cited pages quarterly.
Facts locked in images, PDFs or scripts that crawlers cannot read.
Rewriting other people's data. Summaries are exactly what ChatGPT is citing less.
Over-reading one study. Behaviour shifts with model updates; Ahrefs suggests re-running fan-out analysis each month.
Variations
Price tracker: a living table of what tools in your category cost, refreshed quarterly.
Annual survey, living URL: publish results at /survey and update in place each year.
Compatibility matrix: what works with what, verified by you.
Cost-per-task benchmark: rerun the same AI workflow each quarter and log cost and time.
What to publish if you have no proprietary data
You can create first-party data cheaply. Pick one repeatable test you can run yourself, such as timing the same task in five tools or running the same prompt through several models, and publish the inputs, outputs and your scoring. Survey your own audience with three questions and publish the results in full. Or log your own business numbers, such as what your tool stack costs each month, on a living page.
Sources
Ahrefs: ChatGPT query fan-out study (12 October 2026)
Greenflag Digital: ChatGPT citation study, US finance (21 September 2026)
LLM Pulse: ChatGPT referral traffic fell (13 October 2026)
From today's Daily Briefing → DeepSeek's $0.75 Search API, 71.5% of AI Tokens Go Open-Weight, and How Reddit Lost Its Grip on ChatGPT's Answers