Engine divergence: where the AI engines disagree about citing you

What this check measures

For a single brand and query, we put the identical prompt to each AI engine you have connected and record each engine's own honest verdict: was the brand actually cited in that engine's answer, or not. We then place those verdicts side by side so you can see, at a glance, which engines cite you and which do not. The comparison is built only from engines that were actually measured on that cycle. If an engine was paused, timed out, or returned an error, it is shown as not measured and is never silently treated as a "no." Every per-engine verdict carries the same receipts as a single-engine measurement: whether it was cited, the cited count over the number of successful runs, the published confidence interval around that rate, the stability band, and a one-click link to the verbatim answer. The divergence view adds the cross-engine angle on top of those receipts; it never invents a comparison the underlying measurements do not support.

Why it matters

Teams instinctively treat "AI visibility" as one number, but there is no single AI to be visible in. ChatGPT, Claude, Perplexity, and Google's AI surfaces each draw on different sources and make different citation choices, so a brand can be the top citation in one and entirely absent in another for the very same question. That split is the most actionable signal in the whole product: it tells you which engine is already working for you, which one you are losing, and where the gap is widest. A blended average hides all of it. By showing the per-engine verdicts together, with each engine's own confidence interval intact, you get an honest map of who cites you and who does not, instead of a single score that papers over the disagreement that actually drives your AI traffic.

How we score it

We do not blend the engines into one number. We show each engine's own verdict, computed exactly as it is for a single-engine measurement, and then describe how much the engines agree or disagree in plain language. The cross-engine agreement description is an editorial summary on top of the measurements; the measurements themselves are the receipts, and only the verdicts and their intervals are published. The internal logic that decides how to characterize the spread across engines is deliberately not part of the public page, because it is editorial judgement, not a measurement anyone needs to reproduce. What is reproducible is each engine's underlying citation rate and interval, which are published per engine and can be recomputed from the cited count and the number of successful runs.

Confidence-flag rules

Divergence is only ever computed across engines that were actually measured on the same cycle. An engine that was down, paused, or errored is labelled not measured and is excluded from the contrast entirely, so an outage can never masquerade as "this engine does not cite you." Each per-engine verdict keeps its own confidence interval and stability band, so a wide, uncertain verdict on one engine is shown as exactly that and not hardened into a clean disagreement. Some surfaces are near-deterministic rather than freshly resampled: Google AI Overviews and similar SERP-derived surfaces return what the search result page exposes, so a "miss" there reflects a single source rather than an independent engine making an independent choice. For those single-source surfaces we still show the citation verdict and its interval, but we suppress the cross-engine verdict and label the result single-source, because contrasting a single-source surface against freshly resampled engines would overstate the disagreement.

Common mistakes

• Treating AI visibility as one number. There is no single AI; a brand cited heavily in one engine can be absent in another for the identical question.
• Reading an engine that was down as an engine that does not cite you. An outage is "not measured," never a "no," and we keep the two strictly separate.
• Comparing engines that were not measured on the same cycle. A contrast is only honest between engines that actually answered on the same run.
• Counting a single-source SERP-class surface as an independent disagreement. A SERP miss reflects one source, not an engine independently choosing to omit you, so its cross-engine verdict is suppressed.
• Optimizing toward the blended average. The split between engines is the actionable signal; averaging it away discards the thing worth acting on.

How to fix it

There is nothing to "fix" in the divergence view itself; it is a comparison, and the right move depends on what it shows. When one engine cites you and another does not for the same question, that gap points you at the engine-specific work: the other AI-visibility checks on this report (crawler readability, schema completeness, the source graph) name the concrete levers per engine, and the per-engine playbooks translate the gap into specific moves. When two engines disagree but both verdicts are wide and uncertain, treat that as noise to re-measure rather than a real split to chase. After you act, re-run the measurement and compare each engine's interval against its prior interval, not the headline split, so you can tell a genuine shift from run-to-run variation. Every per-engine verdict links to the exact answer, model, and timestamp behind it, so any disagreement on the page can be opened and verified one click away.

Primary sources

• OpenAI, reproducibility and the seed / temperature parameters (why identical prompts vary)

  OpenAI

• Anthropic, Claude API temperature and sampling

  Anthropic

• Perplexity, Sonar API and online models (independent retrieval per engine)

  Perplexity

• Google Search Central, AI features and your website (AI Overviews draw on Search)

  Google Search Central

Changelog

• 2026-06-23 · Initial publication. Documents the cross-engine divergence view: same question to every connected engine, per-engine honest verdict with its confidence interval, no contrast against engines that were not measured, and suppression of the verdict for single-source SERP-class surfaces.