The Null-Results Registry

The registered predictions

Five entries, none yet reporting. Stated plainly rather than padded: a registry is judged by what it eventually reports, not by how full the table looks at the start.

Not every pre-registered study here predicts a null. The content-freshness and author-credential RCTs register a directional guess instead — a genuine expectation that something will change. Those are not tracked here, but their results are published on the same terms on their own pages and on the studies index.

Why each prediction was made

A prediction is only meaningful if you can see the reasoning behind it. Three of the five are worth setting out.

Schema markup produces no detectable citation change. Registered against the prevailing industry recommendation, deliberately. An observational test across 1,885 pages found citations barely moved, and separate live-fetch tests found major AI systems reading only visible HTML while ignoring JSON-LD. Two independent lines of evidence point the same way, which means a positive result would be a genuine surprise rather than a confirmation.

A majority of cited URLs do not remain cited after 90 days. The reasoning here is weaker than the schema case, and saying so is part of registering it: no prior data exists, because nobody tracks citations forward. It rests on the observation that AI answers are non-deterministic run to run, and that engines already disagree sharply with each other — even two Google surfaces do not agree. If citations turn out to be durable, that is reassuring for anyone investing in GEO, and registering the opposite prediction is what makes the finding credible when reported.

Page speed produces no detectable citation change. The weakest-mechanism entry, and registered partly for that reason. There is no obvious pathway by which load time would affect whether a retrieval system quotes a passage, particularly for cached content. Page speed has a well-documented literature of its own — the Core Web Vitals thresholds and the Web Almanac performance chapter — none of it about citation. It appears on tactic lists anyway, including in the technical audit. Testing something with no plausible mechanism is a reasonable use of a null prediction: a positive result would mean something in the current model of retrieval is wrong.

Note what these have in common. Each predicts against something either widely recommended or intuitively appealing. A registry stocked with predictions that were obviously going to come back null would demonstrate nothing about the process.

What counts as a null result

A study whose registered guess was "no real effect". Once it finishes, either that guess holds up or the study finds a surprise in the other direction. Both outcomes get recorded here, so the registry is not stocked only with the safe, boring nulls that were easy to predict correctly.

Three kinds of null result

"Null result" gets used for three genuinely different situations, and collapsing them is how a weak study gets treated like a strong one.

A well-powered null had enough sample to detect an effect anyone would care about and found nothing; it supports the claim that any real effect is smaller than the threshold tested. An underpowered null would also have missed a moderate effect, and tells you almost nothing — yet it is routinely reported in the same language as the first kind. An inconclusive result is too noisy to support any conclusion, which is not the same as finding no effect, and is the rarest of the three to see published anywhere.

Every entry here will state which of the three it is when it reports.

The file-drawer problem in this field

The metaphor is old and precise. Studies that find nothing go in a drawer. Studies that find something get published. Anyone later surveying the literature sees only the drawer's contents that escaped, and concludes the effect is better established than it is.

In academic fields this has been measured, imperfectly, by comparing registered studies against published ones and counting the gap. That comparison is possible because registration exists. In AI-search research it is not possible, because almost nothing is registered in advance. The size of this field's file drawer is genuinely unknown.

What can be observed is suggestive. Across the tactics tracked on the evidence scoreboard, the overwhelming majority of published findings are positive. Tactics get recommended; almost none get reported as tested and found ineffective. The clearest exception is llms.txt, and it is worth studying as a case: the proposed standard was widely explained and widely adopted before anyone tested it, then a study across 300k domains found no clear effect on citations, most files were never read at all, adoption kept rising anyway, and Google said the idea was purely speculative, repeatedly.

Independent tracking studies and adoption surveys agree; our own read is in does llms.txt work. The nulls arrived, and the recommendation kept circulating regardless. In a field where a dozen widely-repeated tactics have never been controlled-tested at all, a near-total absence of published negative findings is not plausibly because everything works.

There is one prominent exception inside the academic literature, and it is instructive. The one peer-reviewed study in the field included a deliberate negative control, keyword stuffing, and reported that it performed worse than doing nothing. That result survives in circulation mainly because it appeared inside a paper whose other findings were positive. A standalone paper reporting only that keyword stuffing does not work would likely have been much harder to place and much less quoted.

That asymmetry is the whole problem. The result's survival depended on being bundled with good news, not on its own usefulness. This registry exists so that null findings from this site's own studies do not need a positive result to travel alongside them. Which tactics currently have any evidence at all, in either direction, is the job of the scoreboard linked above; the vocabulary is in the glossary, what GEO means and what AI SEO is.

How an entry gets added

The mechanism matters as much as the intention: a registry that depends on someone remembering to add entries will quietly stop being complete. So entries arrive automatically, straight from the pre-registration of any study here whose main hypothesis predicts a null or negative result. No editorial filtering after the fact. The prediction is locked in before collection starts, exactly as written on each study's own page.

That automatic rule is the same one used for crawler-access research, where the pre-registration lives on the crawler guide, the bot registry and the crawler statistics hub rather than here.

How to read a null result correctly

When an entry here reports, the write-up will follow a fixed shape. Knowing that shape in advance makes the eventual results easier to weigh, and it also serves as a template for reading anyone else's null findings.

Two failures overstate a null. Treating "no detectable effect" as "zero effect" is the most common: every study has a detection threshold, and an effect below it is invisible rather than absent. Generalising beyond what was tested is the second — a null on Article schema added to pages that had none says nothing about Product schema.

Two failures understate it. Using a null to dismiss a tactic with other justifications is an error: if content freshness comes back null for citation, that is not an argument for letting pages go stale. And one null does not settle a question; the right response to a single well-run null is to update toward "probably no large effect", not to close the file.

Where this practice comes from

None of this is invented here. Clinical-trial registration exists because trials with unfavourable outcomes were disappearing, leaving a literature that overstated how well treatments worked; psychology's registered reports followed widely-cited findings failing to replicate. In both cases the problem was not fraud but ordinary incentives operating on honest people, and the fix was not asking people to try harder but changing what had to be committed to in advance. AI-search research has the same incentive structure and none of the infrastructure.

That is also why this page exists as a standing list rather than a line in an about page. Anyone can claim to publish inconvenient findings; the claim is only checkable if the findings were promised in advance and are therefore missable when absent.

Limitations

• No null result has been reported here yet. Every entry is a prediction awaiting data. Until at least one reports, this page is a statement of intent with a date on it.
• This registry only covers studies run on this site — it is not a field-wide tracker of null results, and no such tracker exists.
• A registry can be gamed by registering only safe predictions. The check is that the underlying studies are pre-registered publicly, so the full set of predictions is visible, not only the ones that landed here.

Next: see which tactics currently have any evidence at all, in either direction, on the GEO tactic evidence scoreboard. If you are designing your own test, the measurement standard sets out the reporting shape these entries will follow.