Specific facts are what get a page cited by AI answer engines.

Put the specific answer on the page

We took pages that talk around a question and replaced their generic phrasing with the specific figure the question asks for — a price, an interval, a dose — keeping the length within three words. Citation rose on every model we tested, by 37 to 61 points. The same result appeared at 12 questions and again at 48.

Repeating your keyword helps a little, and doesn't get you misquoted

We expected keyword stuffing to do nothing, or to hurt. It helped: repeating a page's topic phrase four to eight times, with no new facts and no extra length, raised citation by +3.8 points [+0.4, +7.5] — roughly 13× smaller than the effect of adding specific figures. No engine went significantly the other way, but on Gemini and Claude the interval includes zero.

We also checked whether stuffed pages get misrepresented more once cited. They don't: the share of attributed claims the page didn't support moved by −0.9 points [−3.8, +1.8], and a second, independent judge model agreed.

Padding a page with filler does not win it citations

Doubling a page's length with filler that states no fact, names no entity and repeats none of the question's words moved its citation rate by +0.4 points [−2.3, +3.1] — an interval comfortably around zero. Length is invisible in this round's variance decomposition too, behind which question was asked, which domain it came from, and the page's format.

But the engines disagree about the sign. On three of the five a padded page is cited less often, each with an interval clear of zero, and two of those hold up after correcting across the panel. The engines that shrug padding off are the two carrying most of the traffic, which is why the pooled figure lands flat. Nothing here rewards padding on any engine; on some it costs you.

A page's share of the answer's sentences — a fairer test of the word-count metrics used elsewhere in this field — did not move either, on any dose.

How the answer is presented does not change whether it is cited

These are the things the optimisation advice actually sells: put the answer first, format it as an FAQ, attribute it to its source, cite where it came from. We applied all four to documents already competing against real alternatives, holding the facts identical — no tactic was allowed to add a figure the original did not have, because a page that says more is already known to win. Pooled across the panel they moved citation by +0.1 points [−3.6, +3.4].

The engines are not unanimous, and the disagreement is worth more than the pooled figure. On one engine these tactics cost a page citations and on another they earn a few, each with an interval clear of zero — but neither survives correction across the panel, and they point opposite ways. Per-cell movement sat at the level you get from running the same unchanged document twice.

Two further probes asked whether the documents were simply too short for order to matter. Lengthening them did not rescue the effect. Rebuilding every document as a full page section did something more awkward: the citation rates split to the extremes, most documents either always cited or never, and the gradeable middle this measurement depends on disappeared.

That middle turns out to be an artefact of cutting the documents short. A short excerpt is a partial answer and earns a partial rate; a full page either contains the answer or it does not. So this null cannot be fixed by running a bigger version of it — the regime it needs is a property of the truncation, not of the content.

Engines largely agree on what deserves a citation

GEO tools are often sold on the idea that every engine is its own optimisation problem. We measured how closely engines agree on which documents they cite, and compared it with how closely one engine agrees with itself when asked the same thing again. Engines agree with each other about as closely as one engine agrees with its own re-run — the disagreement between them is no bigger than ordinary run-to-run noise.

Engines sometimes credit your page with things it doesn't say

Being cited isn't the same as being represented accurately. For every sentence that cited a page, we asked a separate model whether the page actually supports it. Between 13% and 20% of cited sentences made a claim the page doesn't support, depending on which model did the checking. These test pages don't directly answer the question asked — the situation where misattribution is most likely — so pages that do may fare better.

Tracking tools that query the API may undercount your citations

Most citation-tracking tools query engines through their API. We asked the same 12 everyday questions through OpenAI's API and on chatgpt.com without an account. On 3 questions the API cited nothing while the website cited real sources. The reverse never happened, and when both cited, they largely pointed to the same sites.

API: three runs per question, sources pooled. Website: one run. The website may use a different model than the API, so some of the gap could be the model rather than the channel.

The effect holds on questions we have never released

A benchmark that publishes its own corpus is eventually trained on it, and an effect that survives only on questions a model has already seen is not an effect. So a second set of questions was built and deliberately never released — not in this repository, not in any report, not now and not later. Only totals from it are ever published.

Running both sets together, the strongest finding on this page holds on the unreleased half: +47.3 points [+37.5, +57.1], with an interval clear of zero on every engine. Which half a question came from explains one hundredth of one percent of the variance in whether its page gets cited; the change to the page itself explains a quarter.

This is not yet a test for contamination, and the pre-registration says so: at the time of measurement the repository had never been public, so neither set had been exposed and there was no asymmetry to find. What it establishes is the before — a reading that can only be taken once, and that a later re-run is compared against.

One caveat travels with the number. The treatment arm saturates: on most engines every single run cites the page once it carries the figures. So +47.3 is a floor on the effect rather than a measurement of it, and the true size is not identified at this baseline.

The figure above uses a rule corrected after this round was pre-registered. The committed version of the rule dropped any engine that often cited nothing at all, without checking whether that happened only when no page answered the question. One engine was dropped on exactly that basis, and it turned out to show the largest effect here. The pre-registered figure, computed without it, was lower; both numbers and the reason they differ are set out in the report, and the correction now applies to every round.