Answer engines and how to get cited

Why most advice on this is unfalsifiable

Almost everything written about getting cited by ChatGPT, Perplexity or Google's AI answers says the same handful of things: write clearly, answer the question directly, be trustworthy, add structure. None of it is wrong. All of it is impossible to check, which means nobody checks, and a recommendation nobody can fail is a recommendation nobody can follow.

So this guide starts from a narrower question. Not "how do I get cited", which depends on retrieval systems we cannot see and competing pages we do not control, but "what has to be true of a passage before citation is even possible". That question has mechanical answers, and we can run them against our own site and publish what came back.

This measures preconditions, not outcomes. Nobody outside the engines can measure whether they were cited, and any tool claiming to measure it directly is inferring. Passing all four checks does not earn you a citation. Failing them rules one out.

The four preconditions

Answer engines do not quote pages. They quote passages. Everything below follows from that one fact, because a passage gets lifted out of its page and has to survive the trip.

1. Short enough to quote whole. Roughly 50 words. Our working assumption is that a 300 word answer gets summarised rather than quoted, and a summary carries no attribution.
2. Self-contained. It has to make sense with nothing around it. An answer opening with "It depends on your budget" is a fine sentence and a poor citation, because anything lifting it has lifted an orphaned pronoun.
3. Carrying something concrete. A number, a version, a named tool, a price, a date. Generic prose is interchangeable with every other page saying the same thing, which gives nothing a reason to prefer yours.
4. Present in the HTML a crawler receives, with no JavaScript run. This is the one that is invisible in a browser, because your browser runs the JavaScript the crawler does not.

The fourth is the one that silently disqualifies whole sites. If your page renders its content client-side, you can read it perfectly and a crawler can receive an empty shell. It is easy to miss, because your browser runs the JavaScript that hides the problem. You need a tool that does not: curl, view-source, or Search Console's URL inspection.

We ran this against our own site

On 23 August 2026 we checked all 200 answer entries on this site against the first three, and fetched a spread of 12 live pages with a plain GET to check the fourth. The checks run over the opening paragraph of each answer, since that is the passage meant to be liftable. The script is in the repository and the method is short enough to read.

The median opening paragraph is 49 words, with a range of 37 to 63, so the length discipline is holding. Self-containment is nearly universal. And then concreteness collapses: more than half our opening paragraphs contain no number, no version and no named tool.

That is the useful finding, and it is about us rather than about answer engines. We wrote 200 pages to a house style that specified length and directness and did not specify evidence, so we got 200 openings that are well-formed and largely interchangeable. The two rules we wrote down are the two we passed.

The server HTML check came back clean: 12 of 12 pages carried their answer in the raw HTML, with a self-referential canonical, and nothing bailed out to client-side rendering. That is not a virtue, it is a consequence of the site being statically generated at build time. Sites built differently should check rather than assume.

What structured data is currently worth

The standard recommendation is to add FAQPage markup. We did, on every one of those 200 pages, and all 12 sampled pages carry it along with Speakable markup. Here is what that is currently worth.

Google has withdrawn the FAQ rich result. It was narrowed to well-known government and health sites in August 2023, then deprecated, and its documentation was removed outright in June 2026 with the note that the feature is no longer shown in Google Search. The old documentation URL now redirects to that changelog entry. Our own Search Console search-appearance report for the 28 days to 21 August 2026 is empty, which is consistent with it.

This does not make structured data pointless. It labels your content unambiguously, which helps any system trying to work out what a page is, and that job did not go away. But it is not a shortcut, it has never been a ranking factor, and a strategy that consists of adding schema is a strategy with nothing in it.

We are leaving our FAQPage markup in place. It costs nothing to keep, it still describes the page accurately, and the systems that read it are not only Google's. What changed is that we stopped counting it as work.

Quotable does not mean found

There is a comfortable idea that optimising for answer engines routes around the difficulty of ranking. Our own data says otherwise, and it is worth publishing because it is unflattering.

In the 28 days to 21 August 2026 this site drew 885 impressions and zero clicks across all devices. On desktop alone, average position was 54.85. That number is a mean of the topmost position we held on each query, not a place the site sits, and Google's own documentation warns against reading a position value as a page number. What it does tell you is that we are being surfaced far enough down that the click-through rate is near zero however liftable the passage is.

Retrieval has to reach you before quotability matters. For our site the binding constraint is not how quotable the answers are, it is that almost nothing ranks highly enough to be a candidate. Fixing the answers without fixing that would produce very well-formed passages nobody retrieves.

The corollary is where to spend effort. Depth and authority on the handful of topics already close to page one will do more than another hundred well-structured answers, which is the reasoning behind how we now sequence our own writing.

What to actually do

1. Check the fourth precondition first, because it is binary and easy to miss. Fetch your own page with JavaScript disabled, or with curl, and confirm the answer is in what comes back. If it is not, nothing else on this list matters.
2. Put a complete answer in the first paragraph, in about 50 words, that makes sense with the rest of the page deleted. Write it as though it will be read alone, because that is the case you are optimising for.
3. Put something falsifiable in it. A version number, a measured figure, a price with a date, a named tool. If the sentence would be equally true on a competitor's site, it will not be preferred over theirs.
4. Say what your claim does not cover. We think a stated limit is among the clearest signals that a person wrote the page, and it has the advantage of being true regardless.
5. Date anything that moves, and record where it came from. "As of August 2026" plus a source beats a confident undated assertion, which ages into a liability.
6. Then check whether you rank at all for the query. If you sit past position 30, the constraint is retrieval and the fix is depth and links, not phrasing.

The first item is worth doing today even if you do nothing else. It takes a minute and it has a yes or no answer. We have not surveyed how many sites fail it, so we will not put a number on that, but it costs nothing to rule out.

How we keep measuring it

The checks run as a script that writes its own output to a file in the repository. The figures on this page were then transcribed from that file by hand, which is a real weak point and worth naming: the first draft of this page said 44% passed all four preconditions when the recorded run says all three. If this page and the run ever disagree, the run is right.

The script also distinguishes "could not reach the page" from "the page failed the check". That sounds pedantic until it happens: an earlier evidence script here recorded all seven of our security headers as missing because a probe came back empty, and every one of them was present. A proof that cannot tell absence from failure is worse than no proof, because somebody acts on it.

What this does not cover

• This measures preconditions for citation, not citation. We cannot see inside any engine's retrieval, and neither can anyone selling you a tool that says otherwise.
• The concreteness check accepts any digit or capitalised proper noun, so a date passes it. It is a floor, not a quality bar, and 47% is therefore an optimistic reading.
• All three content checks read only the opening paragraph of each answer, not the whole answer. An entry that names a tool in its third paragraph is recorded as failing the concreteness check, and correctly so for our purpose, since the opening is the passage meant to be lifted. It is not a statement about the entry as a whole.
• Average position is a mean of the topmost position we held on each query, across only the queries where we were seen at all. It is not a page number and not a rank the site holds. Impressions and clicks here are across all devices; the 54.85 figure is desktop only.
• The figures on this page were transcribed by hand from the recorded run rather than generated into it. That transcription has already introduced one error, since corrected.
• The self-containment check is a pattern over how an answer opens. It catches an unresolved pronoun. It cannot judge whether the sentence is any good.
• The live probe sampled 12 pages spread deterministically across the corpus rather than all 200, so it would not catch a failure confined to a handful of pages.
• Every figure describes our own site. A site built with client-side rendering, or one already publishing measured content, would produce different numbers and different conclusions.
• Search Console reports what Google saw. It says nothing about ChatGPT, Perplexity, Claude or any other engine, none of which publish comparable data.
• The FAQ rich result status is as Google's documentation stated it on 23 August 2026. This area has changed repeatedly and is likely to change again.

Sources

• Search Central documentation updates: removing the FAQ rich result, Google Search Central. Retrieved 23 August 2026.
• Changes to HowTo and FAQ rich results, Google Search Central Blog. Retrieved 23 August 2026.
• The whatscene.in answer corpus, 200 entries, plus 12 live pages fetched with a plain GET. Word count and opening-pattern checks over the answer data, a concreteness check for a digit or proper noun, and a live fetch with no JavaScript executed and script and style contents stripped before looking for the answer text. Run: scripts/citability.mjs --live, 2026-08-23.
• Google Search Console performance for whatscene.in, 2026-07-25 to 2026-08-21. Search type Web, last 28 days, exported 2026-08-23. Impressions, clicks and average position as reported. Run: Search Console export, 2026-08-23, kept at data/search-console/2026-08-23-performance-28d.zip.

Revisions

• 23 August 2026 First published, with the citability run and the 28 day Search Console export recorded.

This page is revised in place rather than replaced, so its address does not change.