English · Lire en francais

The seven checks that decide if an AI engine cites your page

An answer engine can only cite a page it can fetch, read and trust. Seven checks decide that. Valid JSON-LD. A title built to be quoted. A named author and date. Numbers you can verify. A reachable sitemap and llms.txt. A direct answer in the first paragraph. A robots.txt that lets AI crawlers in.

by Ducens · Published 2026-08-04 · Updated 2026-08-05

Most pages that never get cited are not bad pages. They just never handed the model a clean, readable copy of themselves, so a competitor got quoted instead.

In this guide, you will learn to check the seven things that decide whether an AI engine can cite a page: valid JSON-LD, a title built to be quoted, a named author and date, verifiable numbers, a reachable sitemap and llms.txt, a fast answer up top, and a robots.txt that lets the crawlers in.

Let's go through each one.

1. schema.org JSON-LD, present and valid

JSON-LD isn't decoration. It's the identity card a crawler reads before it reads anything else on the page.

The tag lives inside a <script type="application/ld+json"> block, hidden in the code where a human reader never looks. It names the page's type, headline, author and date in structured data, so the engine doesn't have to guess. Skip it, and the engine reads the visible text alone, more slowly and less reliably.

The failure nobody catches: two plugins each inject their own Article node with a different headline, so the page makes two contradictory claims about itself, and the engine trusts neither. An SEO plugin and a theme both add JSON-LD by default, and nobody turns one off.

To fix it:

1. Open the page source and search for application/ld+json.
2. Count the Article nodes. More than one means two tools are doing the same job.
3. Turn off the extra source, usually a plugin or theme setting.
4. Reload the source and confirm exactly one clean Article node remains.

2. A title, meta and H1 written to be quoted

A title isn't written to rank. It's written to be lifted whole into someone else's answer.

Ranking and quoting are not the same skill. Home | Acme Group ranks fine for the brand name and gets quoted by nobody, because it makes no claim. Elevator maintenance: what the 2026 rules change for public buildings states a claim, and an engine can drop it straight into an answer. The meta description (the summary under a search result) and the H1 (the page's big headline) get the same treatment.

For example, read your title out loud, alone, to someone who has never seen the page. If it doesn't hold up as a sentence, it fails the test.

To fix it:

1. Write the page's single claim in one plain sentence.
2. Put that sentence, or something very close to it, in the title.
3. Repeat a close variant in the H1.
4. Rewrite the meta description as a second sentence that adds detail, not a repeat of the title.

3. A named author and a visible date

An undated page isn't neutral. It's invisible to time, so an engine has no reason to trust it.

E-E-A-T (experience, expertise, authoritativeness, trust) is shorthand for the signals an engine uses to decide whether to believe a page, not just read it. Two of the cheapest signals to hand over are a real name and a date the crawler can find. Skip both, and the page looks the same on day one and five years later, with no reason for an engine to prefer it over a dated competitor.

To be clear: a name in the visible layout isn't enough if it only exists through JavaScript, because a crawler reads the markup (the raw HTML code behind the page) before running any script. For example, a sidebar widget that shows "Published recently" via JavaScript gives a reader a date and gives the crawler nothing.

To fix it:

1. Add a real name, a person or the organisation, directly in the markup.
2. Add the date inside a <time datetime="YYYY-MM-DD"> tag, not only in a design widget.
3. View the page source to confirm both survive with JavaScript turned off.
4. Repeat this on every client page, not just the flagship one.

4. Verifiable numbers, not vague claims

Most companies cites nobody. A number with a scope and a date cites easily.

Of the N pages we audited for this client in June 2026, X blocked GPTBot can be checked, because anyone could ask which pages and get an answer. Engines surface exactly this kind of sentence, because a number is easier to trust and easier to quote than an opinion. The number isn't decoration: it's the hook that gets the sentence picked.

For example, strip every unsourced claim ("most", "many", "often") from a client draft, and it usually loses a third of its length. That's the fluff leaving, not a loss.

To fix it:

1. Find every claim with no number attached.
2. Replace it with your own measurement, scoped and dated, or cut it.
3. Never borrow a round number from someone else's post: measure it yourself, or leave it out.
4. Reread the page once the fluff is gone.

5. llms.txt and sitemap.xml, reachable at the root

A sitemap that lies is worse than no sitemap at all.

Both files live at the root of the domain: yoursite.com/sitemap.xml and yoursite.com/llms.txt. The sitemap lists every real page on the site. llms.txt is a short, plain-text summary for AI systems: what the site is, what it offers. A crawler spends limited time on each site, its crawl budget, so a sitemap full of redirects and 404s burns that budget proving the site unreliable.

For example, a sitemap built once and never touched still lists pages deleted long ago. Each dead link taxes the crawler's patience.

To fix it:

1. Open yoursite.com/sitemap.xml and yoursite.com/llms.txt directly in a browser.
2. Publish whichever one is missing. llms.txt takes about ten minutes to write by hand.
3. Click every link in the sitemap and remove anything that redirects or fails.
4. Recheck both files once a quarter, since sitemaps rot as pages get deleted or moved.

6. An answer in the first paragraph, under 60 words

The top of the page is the only part some readers, human or machine, ever see.

An engine reads the opening first and often stops there. Open with a warm-up about how the industry has changed in recent years, and that's what gets read, then discarded, because it never answers the question. Put the answer first, in under 60 words, and put everything else after it as evidence.

For example, a page that opens "In recent years, the elevator maintenance industry has evolved..." loses the reader and the engine in one sentence. Open with the direct answer instead, and you keep both.

To fix it:

1. Write the direct answer to the page's core question in one paragraph, under 60 words.
2. Put that paragraph first, before any introduction.
3. Move scene-setting or history further down the page, or cut it.
4. Count the words before you publish.

7. A robots.txt that lets the AI crawlers in

This is the single most common cause of zero citations, and the hardest to see: everything else on the page looks fine.

robots.txt is a small text file at the root of the domain that tells crawlers which parts of the site they may visit, and it's the first thing any crawler reads. A security plugin or a well-meaning developer often adds a broad rule that blocks everything by default. Cover an AI crawler by mistake, and the page goes invisible to answer engines while still ranking fine in classic search, because Google's crawler and OpenAI's or Anthropic's are named, and blocked, separately.

Check these agents by name:

To be clear: blocking Google-Extended while keeping Googlebot allowed is a deliberate choice. Blocking it by accident, with every other AI crawler, is the leak that costs you the citation.

To fix it:

1. Open robots.txt and find the line for each crawler above.
2. If one sits under a rule that says Disallow: /, it is fully blocked.
3. Remove or narrow that rule so the crawlers you want can reach the page.
4. Save, reload, and check again. This is the most common cause of all, so check it first, not last.

Key takeaways

• Three of these seven checks are usually already fine on a well-built site. The audit's value isn't finding problems, it's proving which ones are real, in twenty minutes, in writing.
• The deliverable is a page-by-page pass or fail, with the exact line to change and a date on the report.
• That document is what a client pays for, not the twenty minutes behind it.

Charge for the judgement, not the scan.

Our extension, the GEO Site Auditor, automates these seven checks and exports a white-label report, Markdown or PDF, under your own name. Free for five audits a month, 29 dollars once for unlimited use. No subscription, no account, nothing leaves the browser.