Why pre-register
A September 2026 systematic review of 9,249 SEO, AEO and GEO documents found 602 of 769 scored tactics had no controlled test behind them, and that AI visibility had among the thinnest evidence of any topic. It also found 81.9% of published SEO outcomes were positive — which is not a real success rate. It is what gets published.
Pre-registration is the fix for that. The design, the measurement, and the rule for what counts as a result are fixed today. If the outcome embarrasses our own recommendations, it gets published anyway.
The question
FAQPage JSON-LD contributes to AI citation of the page that carries it. Removing it from randomly selected pages will reduce their citation rate relative to matched pages that keep it.
Design
Fourteen pages across three sites we own, each with a visible FAQ and matching FAQPage schema. They are grouped into seven pairs matched by site, format and topic. In each pair, one page was randomly assigned to lose its schema on 13 October. The visible page does not change. Nothing else on any of the fourteen pages changes.
Randomization used a fixed seed — 20260921 — so the assignment can be reproduced and checked. Assignments are final, including where randomization put a page we would rather have protected.
| Pair | Keeps schema | Loses schema |
|---|---|---|
| P1 | Backpacker rap & blerd culture | Afrofuturism lineage |
| P2 | Berserk & grief | Black anime creators directory |
| P3 | The citation oligopoly | How to measure GEO |
| P4 | Schema.org for GEO | Why AI doesn't cite your content |
| P5 | How AI engines decide who to cite | Localized GEO |
| P6 | Agentic commerce | Ranking #1 no longer means cited |
| P7 | Claude Code for enterprise SEO | The 3-hour fix |
Measurement
Forty-two fixed prompts — three per page, natural questions, no brand names — run twice each across ChatGPT, Claude, Perplexity and Google AI Overviews. That is 24 chances per page per wave to be cited. A citation counts only when the page's exact URL appears as a source. Two baseline waves in October, four waves after treatment, outcome measured on the last two.
The decision rule
Per pair, we compare how much each page's citation count changed from baseline. If the page that kept its schema gained relative to the page that lost it, that pair supports the hypothesis. A one-sided sign test across the pairs decides the verdict.
| Condition | Verdict |
|---|---|
| Fewer than 5 pairs differ at all | Inconclusive |
| p ≤ 0.05 | Supports the hypothesis |
| 0.05 < p ≤ 0.10 | Directional only |
| p > 0.10 | No evidence of an effect at this sample size |
There is also a gate. If baseline shows our pages cited on fewer than 10% of opportunities, the treatment is not applied — at that rate the test could not detect anything short of a total effect. We would publish the baseline and redesign instead.
What this test can and cannot find
We simulated the design before collecting anything. The chance it returns "supports" depends on how often the pages are cited to begin with:
| Baseline citation rate | No effect | 25% drop | 50% drop | 100% drop |
|---|---|---|---|---|
| 3% of queries | 2% | 4% | 7% | 23% |
| 10% of queries | 2% | 6% | 16% | 61% |
| 25% of queries | 1% | 11% | 40% | 95% |
Stated now, before the result exists: this test can detect a large effect and cannot reliably detect a small one. A "no evidence" verdict will mean this test did not find an effect — not that FAQPage schema has none.
Limitations
Seven pairs is small, and two are poorly matched. Removing schema may understate an effect if engines retain prior associations. AI responses vary run to run. API output may differ from what a logged-in user sees. The sites are young. And we are not blinded — we own the sites and score the result — which is why the rule and the scoring code are published today.
Materials
prompt-set.csvassignment.jsonscore.py
Changelog
| Date | Change |
|---|---|
| 21 Sep 2026 | Pre-registered. |
| 4 Oct 2026 | Before any data was collected, corrected content on three test pages: removed three passages of unverified first-person detail from one page in P1, and softened three unsupported superlatives across both pages in P5 — on the page and in matching FAQ markup. Pages re-fingerprinted. Assignment, prompts, schedule and decision rule unchanged. Because these edits deployed one day before the first baseline wave, engines may not have recrawled them by then; we note it as a limitation rather than move the schedule. Later the same day we updated the last-modified date on those three pages to reflect the edits. Eleven other test pages also show a last-modified date older than a 21 September markup clean-up; we are leaving those unchanged until the test ends rather than change a freshness signal across the whole test set at baseline. |
How to cite this
If you reference this test before results are in, please say it is pre-registered and in progress.
Summers, J. C. (2026). Test 01: Does FAQPage schema affect AI citation? Pre-registered 21 September 2026. Sable Search. https://sablesearch.com/research/faq-schema-test/
Jason C. Summers
If you run a site and want to replicate this design, the materials above are free to reuse. Replications that disagree with us are the most useful kind.