AI Citation Checklist: 27 Things That Determine If ChatGPT Cites Your Page
Every item here is something we've seen directly affect whether AI engines cite a page. Some are critical — pages that fail these never get cited. Others are incremental. We've ordered them by how often they're the root cause of the problem.
Short answer: the 27 items break into seven areas an engine checks in turn: whether a crawler can reach the page, whether the content is organized for extraction, whether the vocabulary matches how buyers search, whether claims are sourced, whether the markup is machine-readable, whether other sites vouch for you, and whether the page is current. The critical and high impact items account for the bulk of citation variance, so fix those before the incremental ones.
Go through this on your key pages, or request a visibility assessment and we'll score all 27 for you.
Everything a website needs in order to be cited in AI answers sits in one of seven areas, and the table below is the whole checklist compressed into them. The order is the order an engine runs into them, which is also the order to fix them in.
| Area | What the engine is deciding | The fastest test on one page | Items |
|---|---|---|---|
| Technical | Whether it can fetch and read the page at all | View source and search it for a sentence you can see on the screen | 7 |
| Structure (12%) | Whether it can lift a clean answer without carrying the rest of the page | Read the first sentence under each heading on its own | 5 |
| Alignment (35%) | Whether the page is about the question that was actually asked | Compare your headings against the words a buyer would type | 4 |
| Grounding (13%) | Whether the claims on the page can be checked | Find every number and look for its source in the same paragraph | 3 |
| Schema | Whether the markup agrees with the visible page | Run the page through a structured data validator | 3 |
| Substantiation (15%) | Whether anybody else says the same thing about you | Search your company name and read what third parties state | 3 |
| Newness (5%) | Whether the page is current enough to be worth quoting | Check that the visible date, the schema date and the last real edit agree | 2 |
The order matters more than the count. An engine that cannot fetch the page never reaches the vocabulary question, so one failed technical item takes the other 26 with it, and no other area behaves that way. The percentages are the weights the SIGNALS framework gives each dimension, and the two areas without one are prerequisites rather than scored dimensions.
Technical Can AI engines actually reach your page?
An AI engine can reach your page when the content arrives inside the HTML the server returns, no crawler rule blocks it in robots.txt or at the edge, and nothing sits behind a login. Failures here are silent, because nothing in your analytics records a fetch that never happened, and they cost you every engine at once rather than one of them.
Structure (12%) Is content organized for extraction?
Content is organized for extraction when a question sits in the heading and the answer sits in the first sentence underneath, in a sentence that still makes sense once it is lifted away from the page. An engine quotes passages rather than documents, so the unit of work is the passage.
Alignment (35%) Does your vocabulary match buyer searches?
Your vocabulary matches buyer searches when the page uses the words a buyer puts into the prompt rather than the words your product team uses internally. This is the heaviest dimension in the SIGNALS framework at 35%, and it is the only page-level lever that has measured large in published work, which is why the four items below are worth more than the twenty-three others put together.
Grounding (13%) Are claims sourced and verifiable?
A claim is verifiable when its source sits in the same paragraph as the number, close enough that an engine lifting the sentence carries the attribution with it. A figure whose source is two sections away reads as unsupported to a machine that only takes the passage, which is the whole reason this dimension is cheap to win.
Schema Is content machine-readable?
Markup is machine-readable when the JSON-LD states the same facts the visible page states, and that agreement is the only version of it that is safe to ship. Schema does not win a citation on its own. Its job is to remove ambiguity about what the page is, who published it and when it last changed.
Substantiation (15%) Do others vouch for you?
Others vouch for you when the same facts about your company appear on sites you do not control, which is what an engine checks a claim against before it repeats it. This is the half of the work that does not happen on your own pages, and it is the slower half, so start it early rather than last.
Newness (5%) Is the content current?
A page reads as current when the visible date, the schema date and the last real edit all agree, and when that edit changed the substance rather than the date. Engines disagree about how much they prefer recent pages: Discovered Labs put the median age of a page cited by Claude at about five months, against eight months and beyond for ChatGPT and Gemini, in the same analysis of roughly 2 million citations across 10,000 cited pages (Discovered Labs, What actually drives AI citations, last updated August 2026, read at source 23 October 2026; the firm sells AEO services, so what it chose to measure is not disinterested). Calibrate the cadence to the most demanding engine, and move the date only when the content moved with it.
How should you actually use this checklist?
Use this checklist in three passes rather than in one sweep, because the seven areas have different scopes and only one of them is site-wide. Start with your homepage, your pricing page and whatever page you most want cited when somebody asks about your category. Those are the three pages that matter most for AI visibility.
Don't treat this as a complete-everything list. The Critical and High impact items are where your time is best spent — they account for the bulk of citation variance. Lower impact items are real but incremental. Fix the big stuff first and re-audit to see the score move.
Run it as three passes rather than as one sweep, because the areas have different scopes. The technical seven are a site-level pass and you do them once: check robots.txt and the edge rules, confirm the pages render server-side, and confirm nothing important needs a login. The alignment and structure items are a per-page pass and they are where the time goes, since rewriting a heading into the question a buyer asks means knowing which question that is. Grounding, schema and newness are a finishing pass on the same page, and they are quick once the first two are done.
Then wait before you re-score. A page you fixed today is still being answered from whatever the engine last fetched, so a re-check the same week measures the old page, and how long it takes ChatGPT to index a new page sets out the lag to expect. Re-run the same prompts you ran before the edit, worded identically, or the comparison measures your wording rather than your page.
If you'd rather not do this yourself, a SIGNALS visibility assessment scores all 27 signals, tells you exactly what to fix and in what order, and we do the work.
Frequently asked questions
What is an AI citation checklist?
An AI citation checklist is a structured list of page-level signals that predict whether AI engines like ChatGPT, Claude, and Perplexity will cite a web page. It covers technical accessibility, content structure, vocabulary alignment, credibility signals, and schema markup.
What is the most important factor for AI citation?
Vocabulary alignment, meaning how closely your page language matches the way buyers prompt an engine, carries 35% of the citation likelihood weight in the SIGNALS framework and is the strongest page-level factor measured at scale. Discovered Labs (2026) analysed roughly 2 million AI citations over six months across 10,000 cited pages and put the standardised effect of prompt-content alignment at +0.37, with a 95% confidence interval of +0.33 to +0.41, about three times the next-strongest page-level signal, and +0.35 after its double machine learning check removed confounder variance (read at source 23 October 2026; Discovered Labs sells AEO services, so the choice of what to measure is not disinterested). The same study found a domain's AI-perceived authority six times more influential than the strongest page-level feature, which is the honest limit on what any page-level checklist can do for you.
How many of these items do I need to pass?
Pages scoring 70+ on the SIGNALS framework typically pass 18–22 of 27 items. The four highest-impact items alone — buyer vocabulary in headings, sourced statistics, FAQ with JSON-LD schema, and server-rendered HTML — account for over 60% of citation likelihood variance.
Can SIGNALS check all 27 for me?
Yes — a SIGNALS visibility assessment scores all 7 dimension categories (which cover all 27 checklist items) for you, and shows exactly where each page falls short across the four AI engines.
Related reading
See where your company stands across all four engines.
A free, PULSE-powered visibility assessment maps exactly where you’re cited and where you’re invisible — against the competitors winning your category, category by category.
Request a free visibility assessment →