Most advice about writing for answer engines is assertion. This page keeps to what has been measured, names who measured it, and marks the places where the honest answer is that nobody knows yet. It is about the page itself: what goes on it, how it is built, and which of the popular moves show no effect at all.
On this page
- The thing that gets cited is a passage
- Structure the evidence supports
- Moves that measure as nothing
- Seven questions to ask any number
- Freshness is about the claim, not the stamp
- Ranking still matters, differently
- Written with AI, and what that costs
- When the answer says nothing about you
- The rules we gave ourselves, including the expensive ones
Key takeaways
- Answer engines quote passages, not pages. In one coded sample, cited passages carried a 2025–2026 date 80% of the time against 53% for passages shown but not cited.
- The structural signals that travel with citations are unglamorous: a table of contents in the HTML, Article and FAQPage markup, five or more H2s, a table, a thousand words. Only 14% of pages have a table of contents.
- llms.txt showed no measurable effect in a study of 96,418 URLs. Breadcrumb, HowTo and QAPage markup showed none either.
- 43% of AI Overview citations went to pages in the organic top 10. The other 57% came from below it or from pages not ranked at all.
- 75% of cited pages had been updated within a year — but more than a quarter of those “fresh” pages were first published over two years earlier. Freshness is manufactured by updating, not by publishing.
- Every figure above is one vendor’s measurement of one question set. The seven-point check below is how we decide which ones to repeat.
The thing that gets cited is a passage
An answer engine does not cite your article. It lifts a span of text and attributes that, or uses it and attributes nothing. The difference between those two outcomes turns out to be visible in the text itself.
Bart Magera coded 112 passages from 40 SEO queries and found that the passages engines linked to were more often dated and more often named their subject outright. Cited passages showed a 2025 or 2026 date 80% of the time; passages that Google displayed but did not credit, 53%. The sample is small enough that the uncited group rests on a handful of passages, and we said so when we wrote it up. The direction is still the most actionable finding we have read all year.
What it changes in practice: write the sentence that carries a claim so it can stand alone. Who measured it, when, how large, what it says. A paragraph that depends on the three paragraphs above it to make sense is a paragraph that will be absorbed without a link.
Structure the evidence supports
The largest study we have read on this paired 48,527 cited pages with a non-cited page from the same domain, so brand strength drops out of the comparison. Its findings are correlations, not proven causes, and they are strikingly dull: a visible table of contents travels with 1.7 times the citation rate, Article markup 1.6, five or more H2 headings 1.6, FAQPage markup 1.5, a table with a header row 1.5, a thousand words or more 1.5. Pages carrying many of these are cited about 2.1 times as often as barely built ones.
Two details matter more than the list. The first is that only 14% of pages have a table of contents, which makes it cheap ground. The second is that a table of contents has to be in the HTML; if a script assembles it after load, a crawler that does not run scripts never sees it.
None of this is a content strategy. It is template work: build it once and every page inherits it. We ran the whole checklist against our own archive and published the scorecard, failures included. How the markup fits together — rather than which types are present — is a separate question we measured across twenty sites in connected schema or flat tags.
Moves that measure as nothing
The null results deserve equal billing, because they are where budgets go to die.
In that same dataset of 96,418 URLs, llms.txt showed no measurable advantage. Breadcrumb, HowTo and QAPage markup showed none either. Recipe schema was associated with a tenth of the citation chance, which is a statement about recipe sites rather than about the markup. Keep breadcrumbs — they serve readers and the results page — but do not count them as progress on this front.
The largest single figure in the study, the current year in the page title at 2.7 times, is also its least certain: the authors added the feature after the analysis was planned and say plainly that the year most likely stands in for content that was genuinely refreshed. Writing 2026 into a headline about a 2026 dataset is accurate. Writing it into everything is the oldest trick in the catalogue, and it is the kind of claim we took apart in three GEO claims tested against evidence.
Seven questions to ask any number
Everything on this page comes from vendor research, because that is nearly all there is. We apply the same seven questions to each study before quoting it, and we publish how many it clears.
- Which engine, and which model version.
- Collected through the product’s own interface, or through an API — the two can answer the same prompt differently.
- What dates the collection covers.
- Which region or market.
- The prompt set, published or not.
- How many repeats.
- Whether run-to-run variability was measured or ignored.
A study clearing three of seven is still worth reading; it is not worth forecasting with. The reason to be strict is arithmetic rather than suspicion: in one study’s own sample-size test, a feature looked 23% less common on cited pages at 833 URLs, 31% more common at 4,600, and 19% more common at 96,418. The direction only settled above roughly 5,000 URLs. A GEO study built on a few hundred pages can land on either side of zero by chance.
The same logic applies to tracking your own visibility, where spending a query budget on more prompts beats repeating one prompt: twenty prompts beat twenty days of one.
Freshness is about the claim, not the stamp
Ahrefs measured 47,097 citations from ChatGPT, Gemini and Perplexity against the dates on the pages behind them. Three quarters of cited pages had been updated within the past year, 88% within two. The finding underneath that one is the useful one: more than a quarter of those “fresh” pages were first published over two years earlier. Freshness in this data is manufactured by updating existing pages, not by publishing new ones.
Which puts a condition on it that no dataset can enforce. An update that changes a date and nothing else is a date change. We learned that from the wrong side: a cosmetic pass across our own archive stamped twenty-six articles with the same update date inside two minutes, and we had to reconstruct the real dates from revision history. The rule we now hold ourselves to is that the date moves when a reader would get something new from reading again — never for a renamed heading or a swapped phrase. The mechanics of that are in the technical pillar; the full picture of what recency does and does not buy is in content recency in AI search.
Ranking still matters, differently
Conductor matched 167.9 million AI Overview citations against Google’s organic results for US searches between 15 June and 15 August 2026. A page ranked first was cited by that search’s AI Overview about one time in four — 24.9%. At position ten, 10.2%. Position still helps, and it helps steadily rather than dramatically.
The number that reorders priorities is the other one: 43% of citations went to pages in the organic top 10, and 57% came from below it or from pages not ranked for that query at all. Positions one to three together supplied just over a fifth of all citations. A page that cannot realistically reach the top three can still be cited, which is a different strategic position from the one most ranking advice assumes. We set it out in most AI Overview citations come from outside the top 10.
Whether a citation is worth having is a separate question again, and the answer is less comfortable: citations and traffic parted company in the 2026 data.
Written with AI, and what that costs
We write with AI and say so, which puts us in a position to test the claims about it rather than repeat them. We ran a detector across our own first twelve articles, nine of which were AI-written, and published what came back, including the parts that did not flatter us.
Two things we would tell anyone doing the same. Detector scores are not a quality measure and not a ranking factor; what they catch is uniformity, and uniformity is worth removing for its own sake. And the tells that matter are not exotic — repeated heading patterns across an archive, the same stock transitions, every article opening the same way. We had all three and had to strip them out by hand.
On the surrounding questions — whether provenance marks in generated text affect ranking, and which year-on-year shifts in how marketers use AI survive scrutiny — we have what is actually known about text watermarks and which survey changes hold up.
When the answer says nothing about you
The common failure is not being described badly. It is not being mentioned. Asked about a category, a model names the brands it has material about and passes over the rest in silence — no error, no gap the user can see, nothing to correct. We wrote that up as silence, not error.
This changes what content is for. A page that exists so a model has something specific to say about you — what you do, for whom, with what evidence — is worth more than another piece competing for a head term. And it has to be legible to a machine that reads a passage out of context, which is where the first section of this page comes back round.
Before concluding you are absent, check against the noise. Single observations of these systems are close to worthless, and the mechanism behind which sources get selected is set out in how answer engines cite.
The rules we gave ourselves, including the expensive ones
Rules are easier to judge when someone states what theirs cost them. Ours, and the price:
Name the publisher, the author and the date in the sentence carrying the claim. This is the passage-level finding applied as a habit. It makes sentences longer and it is the single thing we would keep if we had to drop everything else.
State how many method points a study clears before quoting its numbers. Costs a paragraph in every piece. Buys the ability to disagree with a vendor without hand-waving.
Publish the measurements that go against us. Our own archive fails three of the eleven citation signals, our index coverage is four pages of forty-one, and both are published. A site that only reports its wins is a site whose numbers cannot be checked.
Few outbound links, by choice. Ten or more external links is a 1.3x signal in the study above, and we clear it on no page at all: our rule is to name sources in the prose instead. We think attribution a reader can see beats attribution a crawler can count. It is a judgement call, we may be wrong, and we are not going to pretend it is free.
Correct in public. When a source retracts something we repeated, the text changes and the change is logged. Writing quickly about moving subjects guarantees this happens; the only choice is whether anyone can tell.
FAQ
Is llms.txt worth publishing?
On the evidence, no — it showed no measurable effect across 96,418 URLs. It costs almost nothing to serve, so publishing one is not a mistake; counting it as work done is. If you have a choice between llms.txt and putting a table of contents in your HTML, the table of contents is the one with a measured association.
How long should a page be?
A thousand words or more carries a 1.5x association, which the study’s authors attribute to longer pages carrying more structure rather than to length itself. Treat it as a floor below which there is probably nothing to structure, not as a target. Padding a 400-word answer to 1,200 gives you a longer 400-word answer.
Should I update old pages or publish new ones?
The citation data leans towards updating: more than a quarter of pages counted as fresh were first published over two years earlier. The condition is that the update has to change what the page says. A refreshed date on unchanged text is not freshness, and at scale it teaches engines to disregard your dates.
Does any of this work for a small site?
We cannot tell you from our own data, and we will not pretend otherwise. This site had four pages indexed of forty-one submitted when we last measured, which is below the level at which a 1.3x factor could become visible. The structural work is cheap and survives being wrong; the strategic work — being worth linking to — is the part that gates everything else.
