There is no shortage of data about which websites get cited in AI answers. Ahrefs runs an ongoing citation tracker through its Brand Radar tool. 5W Research has synthesized nine separately published datasets into a single estimate of ChatGPT's citation mix. Similarweb, Semrush, Profound, and several other vendors publish their own numbers on a rolling basis. What's harder to find is agreement between any of them. Depending on which study you read, Reddit's share of ChatGPT citations is somewhere between roughly 12% and roughly 17%, collected over different windows with different query sets and different definitions of what counts as a citation.
We're not publishing this to criticize any of that work — most of it is more rigorous than the average trade-press citation stat, and we cite it regularly. We're publishing it because the disagreement itself is the story: citation share is sensitive to methodology in ways that make single-vendor numbers hard to compare, and the only way to know what's driving a given result is to run the study yourself and show your work. That's what our study program is for.
What we're testing
Five studies are in progress or queued, each addressing a distinct question rather than repeating what's already been measured elsewhere:
A large-sample citation audit of ChatGPT, designed at a scale (50,000 queries) and with a documented methodology that lets readers see exactly what's driving our numbers relative to existing estimates.
A matched, same-query, same-day comparison of ChatGPT, Gemini, and Perplexity, since single-engine studies can't show how citation behavior diverges across engines for identical queries.
A controlled before/after test isolating schema markup as a single variable, because AEO tool vendors routinely recommend it and we have not found a published controlled test of the claim.
A small, pre-registered experiment applying a realistic bundle of AEO edits to 20 real pages, tracking citation rate change over a defined window.
A composite AI Visibility Index combining citation and mention data across four engines into a single, transparently weighted score, intended to give a defensible ranking where single-engine leaderboards disagree.
Each of those studies has a full methodology written before any data is collected, and none of them has results yet — because the results don't exist yet. That's deliberate, and it's the core of how this program works.
Why pre-registration
Academic research has a well-documented problem with results-first analysis: when a researcher can choose their methodology after seeing the data, it's easy to arrive, even unintentionally, at whatever conclusion looks most interesting. The 2023 Princeton and IIT Delhi paper that introduced the term "Generative Engine Optimization" ran its tests against a fixed benchmark specifically so the evaluation couldn't be adjusted after the fact. We're borrowing that discipline for a much less academic purpose.
Concretely, that means: before we collect data, we publish the query set (or the method for constructing it), the engines and access configuration, the date range, how we'll parse and classify citations, and the statistical approach we'll use to report results. Readers can see all of that before a single result is in, and can hold us to it. If our methodology changes mid-study, we'll say so and explain why, rather than quietly revising the plan to match what we found.
How we'll publish data and corrections
When a study's results are ready, we will update its existing page rather than publish a new one, so the methodology and results live together and the updated_date reflects when the results landed. Alongside the write-up, we intend to publish the underlying data: query lists, raw citation extractions, and the classification rules used to turn a raw AI response into a "citation" or "mention" count. Readers and other researchers should be able to check our work, not just our conclusions.
If we get something wrong — a miscounted query set, a misclassified domain, a statistical error — we will correct it visibly on the affected piece, noting what changed and when, rather than silently editing the page. Citation studies in this field move fast and few of them are error-free on a first pass; we'd rather be transparently corrected than quietly right.
What this program won't do
We won't present a single snapshot as a permanent ranking. 5W Research's synthesis noted that Reddit's ChatGPT citation share swung from roughly 60% to roughly 10% of prompt responses within a two-week span in September 2025 — a reminder that any citation study, including ours, describes a moment, not a fixed state. Where it matters, we intend to re-run studies periodically rather than treat one collection window as the final word.
We also won't publish results before they exist. Every study in the program is being prepared with a complete methodology and clearly marked placeholders instead of numbers, because we have not run the studies yet. Each one will be published here, methodology and results together, once its data has been collected and checked — follow this page or subscribe to the newsletter for updates as each study lands.
