How to get cited by ChatGPT: 9 tactics, every one graded

1. The answer, and the grades

Getting cited by ChatGPT is a specific case of the broader problem covered in what generative engine optimisation is, and it is not the same problem as ranking on Google. Only about 38% of AI-Overview citations come from pages that rank in the top 10 for the query, per Ahrefs, 863k keywords, and that share is still falling. Extractability now matters as much as authority.

What makes this guide different from every other list of nine tactics is the last column. Each step below carries an evidence grade, so you can see which rest on a controlled experiment and which rest on a correlation somebody published once.

Read that table as a sequencing tool, not a ranking. Bing indexing goes first because it costs nothing and gates everything. The three tested tactics go next because they are the only ones with a controlled result behind them. The rest are worth doing on judgement — just don't report them as measured levers. The same grading discipline, applied across the whole field, is in the GEO tactic evidence scoreboard.

2. The three tested tactics

Evidence These come from one peer-reviewed controlled study: the 2024 Princeton "GEO" paper, which tested isolated content changes across roughly 10,000 queries. The full paper and its abstract page carry the method.

Cite credible sources. Counter-intuitively, linking out to authoritative third parties raised a page's own AI visibility by up to 40% — the biggest single lever they tested. Cite primary sources (studies, official docs, first-party data) over blog roundups, and link the specific claim rather than a homepage. Original research is the strongest version of this, and the most-cited domains study shows how much of the citation pool primary sources take.

Add named quotations. Quotation addition produced a +37% lift in the same study, and quotes double as first-hand experience signals. A sentence like "'ChatGPT rewards specificity over polish,' says [expert]" gives the model a clean, attributable unit to lift. The failure mode is inventing or paraphrasing a quote, which is worse than having none.

Add specific statistics. A +22% lift. Replace "AI Overviews significantly reduce clicks" with "AI Overviews cut clicks to the #1 organic result by ~58% (Ahrefs, 300k keywords, Dec 2025)." Number, metric, population, date, source — that five-part unit is exactly what an AI answer needs to quote you confidently. The underlying click-loss study is the traceable version of that claim, and the provenance audit shows how rarely a circulating figure survives that trace.

The pages that get cited read less like brochures and more like briefings — quotable, sourced, and confident.

One honest caveat before you spend a quarter on this. Those magnitudes come from one study, on one set of queries, at one moment in these systems' history. Nobody has publicly replicated them. The mechanism — specific, sourced language is easier to quote — is plausible on priors. The percentages should be held loosely.

3. The plumbing: get indexed by Bing

Fact ChatGPT's web search runs on Bing's index — if your pages are not in Bing, ChatGPT cannot cite them. This is the most-skipped step in every GEO checklist and the only one on this page that is a mechanism rather than a measured effect. 87% of ChatGPT search citations match Bing's top organic results, versus 56% for Google. Google Search Console will not help you here; Microsoft's own side-by-side comparison explains what Bing Webmaster Tools reports that GSC does not.

1. Create a Bing Webmaster Tools account and build a valid sitemap to submit. Bing has published its own note on why sitemaps still matter in AI-powered search, and accepts up to 10,000 URLs a day.
2. Turn on IndexNow so new and updated URLs are pushed to Bing within minutes — the protocol is now widely adopted.
3. Verify: run a site:yourdomain.com search on Bing and confirm your key pages appear.
4. Verify: check that OAI-SearchBot is not blocked in robots.txt — see GPTBot vs OAI-SearchBot for the exact rules and how to confirm them in your logs.

If you do nothing else on this list, do this. It has its own guide: Bing Copilot SEO. Google's surfaces work differently again — see AI Overview statistics and the delta between AI Mode and AI Overviews.

4. The five untested tactics

Everything below is worth doing. None of it has a controlled test behind it, and the grade on each says why.

Answer the question first. Hypothesis Put the complete answer in your first two sentences, before any preamble. Citations are widely reported to skew toward the opening of a document. The specific percentages in circulation could not be traced to a disclosed method, so none is quoted here — treat the direction as the claim, not a figure. Apply the same rule inside the page: open every H2 with a 40–60 word answer to the question that heading implies, then expand with 200–400 words of support. The same principle underpins practitioner guidance on LLM-friendly content. The risk is over-applying it into a summary with padding attached.

Structure for extraction. Hypothesis Headings, short paragraphs, lists and tables give a model clean, bounded units to grab. Comparison tables are widely reported to be cited more often than equivalent prose, though the multipliers in circulation trace to no disclosed method. Question-style H2s map to how people phrase AI queries. Keep paragraphs to two to four lines; add a table whenever you compare, a numbered list whenever you describe a process. If those units are assembled by JavaScript, none of it counts — rendering is expensive even for Google, and whether AI crawlers pay that cost is an open experiment.

Keep it fresh. Hypothesis One analysis reported that 76% of ChatGPT's most-cited pages had been updated within the previous 30 days, and that cited pages skew around 25% newer than typical organic results. That is an association between two things measured at once, not a demonstration that updating a page causes citation. Put your most important pages on a quarterly refresh cadence and change the content, not just the timestamp. No controlled test isolates freshness — it is graded a hypothesis in the dedicated study page, and how fast a refresh even reaches an engine is the crawl-to-citation latency question.

Freshness is also the most plausible reading of the biggest shift in AI search: the share of AI Overview citations coming from Google's top 10 fell sharply in under a year, which is what makes room for challengers.

Earn brand mentions. Hypothesis One analysis reported branded web mentions as the strongest correlate of AI Overview citation (0.664 correlation) and roughly 3× more predictive of AI visibility than traditional backlinks — a pattern that also appears in correlational ranking-factor work. The direction is unestablished: sites that get mentioned a lot are also, generally, good sites. Treated as causal, this is the most over-sold claim in the field; see brand mentions versus backlinks. The practical version — get quoted in roundups, contribute real expertise on forums, go on a podcast — is cheap and low-regret regardless.

Build author E-E-A-T. Open question The weakest-evidenced tactic here: no test shows that adding a byline changes citation rate. The reasoning is sound and the cost is low. Give every article a visible byline linking to a real author page; give that page Person schema with a stable @id and a full sameAs array, and reference the same @id on every post. On the page itself: full name, photo, role, years of experience, named areas of expertise, and first-hand proof. The author E-E-A-T study page sets out what would actually test it; YMYL topics are where it plausibly matters most.

5. How ChatGPT picks a source

When ChatGPT answers with sources it is not reasoning from training data. It runs a live search on Bing's index, retrieves candidate pages, and chooses which to quote. Vendor teardowns of how ChatGPT chooses its sources describe a consistent funnel, and it explains the whole priority order above.

You cannot manufacture domain authority overnight, but you can control freshness and extractability immediately — which is why a lower-authority page sometimes beats a big brand whose page is stale or buries its answer. Being retrieved is not enough; you have to be the most quotable of the candidates.

6. How to measure whether it worked

AI citations do not appear in Google Search Console, so you need a deliberate routine. Combine three signals, monthly rather than daily, because AI answers jitter from run to run.

• A fixed prompt panel. Write down 15–30 questions your reader would ask, run them through ChatGPT, Perplexity, Gemini and AI Overviews on a fixed cadence, and log which sources get cited. Freeze the wording — this is the ground truth, and the same fixed-panel method behind our AI Citation Tracker.
• An AI-visibility tool. Profound, Otterly, Semrush's AI visibility features and Ahrefs' Brand Radar automate a version of the panel at scale.
• Server-log crawler hits. Filter for OAI-SearchBot and GPTBot using the exact strings in OpenAI's bot reference and the AI Bot Registry. Crawling precedes citation, so rising crawler activity is the earliest leading indicator.

Five mistakes produce nearly all the fake wins in this field.

• An unstable prompt panel. Reworded prompts move your numbers for mechanical reasons.
• A single run per prompt. Report how often you appeared, not whether you appeared.
• Personalisation. A logged-in account that remembers your brand is not a neutral observer.
• Counting crawls as citations. Being fetched is a prerequisite, not an outcome.
• No control. A comparable page you deliberately leave untouched is the cheapest control available, and almost nobody runs one.

Measure each engine separately: only ~11% of domains are cited by both ChatGPT and Perplexity. The full procedure is the AI visibility measurement standard.

7. Why pages stay uncited

Most uncited pages are not bad. They are making one of a handful of avoidable errors, and the first two account for most of them.

The structural ones are catchable with the AI Overview Exposure Checker and the technical GEO audit. The buried-answer failure is the one worth seeing side by side. Both openings below carry the same facts.

Same facts, opposite outcomes. The rewrite is editorial, not technical, and it is the cheapest change on this page.

8. What does not transfer between engines

People say "AI search" as if it were one system. It is several, and a tactic that works on one can do nothing on another.

Hypothesis The pattern is that page-level tactics generalise and plumbing-level tactics do not. That is a reasonable reading of the evidence, not a tested claim. Apply the page-level tactics everywhere; treat the plumbing as a per-engine checklist. Perplexity, Claude, Gemini and the concordance study cover how far apart they sit.

9. What we still do not know

Open question The gaps here are larger than the settled parts.

• Whether the tested magnitudes still hold. The Princeton figures come from one study, one query set, one moment. Nobody has publicly replicated them, and the vendor studies that exist, such as Semrush's AI Overviews work, measure a different surface.
• How the levers interact. Effects are usually measured one at a time and then quietly summed. Does adding statistics help as much on a page that already cites sources?
• The direction of the brand-mention relationship. The correlation is well documented and equally consistent with mentions being a symptom rather than a cause.
• How quickly any of this decays. If an engine changes its retrieval stack next quarter, which of these nine survives? No longitudinal answer has been published.

That is not a reason to do nothing. It is a reason to hold the specifics loosely, run your own panel, and treat every number here as a snapshot rather than a law.