Why two AEO platforms can disagree on the same brand's citation share — and what to do about it

Questions this guide answers

• Why do AEO platforms report different citation shares?
• How is AEO citation share measured?
• What questions to ask an AEO platform about measurement?
• Is AEO measurement reliable?

Direct answer

AEO citation share is measured by running a set of prompts against AI answer engines on some cadence, parsing the engines' outputs for citations, and aggregating the results. Each of those four steps — prompt-set design, cadence, parsing, aggregation — involves choices that materially change the final number. Two AEO platforms making different choices will produce different numbers for the same brand, and neither has to be wrong.

What this essay is not: a claim that any specific platform's number is incorrect. What it is: an engineering note on the four sources of disagreement, and the five questions that surface those choices in a vendor conversation. SolCrys answers all five at the end, with explicit uncertainty disclosures.

AEO measurement is survey design, not crawl analytics

The category's marketing language often borrows from search analytics — "share of voice," "visibility," "impressions." Those words come from a world where measurement is deterministic: a crawler visits a page, a server log records the request, the numbers are facts about events that happened.

AEO measurement is not that world. There is no comprehensive log of when an engine cited a brand; the engine doesn't expose one. What an AEO platform does is sample: it queries a chosen set of prompts at a chosen cadence and counts what comes back. The output is a statistical estimate, not a comprehensive count.

The right analogy is Nielsen TV ratings or a political poll, not Google Analytics. The survey design — which prompts, how often, parsed how — determines the output. Different surveys, different outputs. Disagreement is structural, not a bug.

Source of disagreement 1: prompt-set composition

The single biggest driver of disagreement is which prompts the two platforms track. If platform A is tracking 200 generic category prompts and platform B is tracking 800 prompts including long-tail intent, their measured citation shares for the same brand will differ — sometimes substantially. Neither is wrong; they're measuring different populations.

Within the prompt set, three sub-choices matter: where the prompts come from (search-volume data, community questions, engine follow-ups, brand-supplied, synthetic), how often the prompt set is refreshed, and whether the prompts target the brand's actual buyers or generic category language. SolCrys's Golden Prompt Set methodology is our specific design choice on this; other platforms make other choices.

The question that surfaces it: how many prompts does the platform track for my category, where did they come from, and how often is the set refreshed?

Source of disagreement 2: sampling cadence

Two platforms running the same prompt set at different cadences will see different citation patterns. If platform A samples each prompt once a week and platform B samples once a day, B captures more of the within-week variance — including engine update effects, time-of-day differences, and short-lived citations that A misses entirely.

Cadence also affects how the platform handles freshness. A daily-sample platform can show a citation that appeared on Monday and disappeared on Wednesday; a weekly-sample platform may never see it. Neither is more accurate; they're answering different questions.

The question that surfaces it: how often is each prompt re-queried, and how is engine-side intra-day variance handled in the reported number?

Source of disagreement 3: engine non-determinism

AI engines are non-deterministic in a way classic search engines aren't. The same prompt, asked of the same engine, at the same minute, twice in a row, can produce different citation sets. Engines vary their outputs through temperature settings, retrieval re-ranking, and other internal choices that AEO platforms can't directly control or fully observe.

How a platform handles this matters. Some platforms run multiple samples per prompt and average; some run a single sample and accept the variance; some run a single sample but flag low-confidence outputs. Each choice changes the reported number.

Worth being explicit: this isn't a complaint about engines. Non-determinism is a feature of how generative AI works, and it's not going away. The question for an AEO platform is what it does with the non-determinism, not whether it can eliminate it.

The question that surfaces it: how does the platform handle the case where the same prompt produces different outputs across re-runs?

Source of disagreement 4: what counts as a citation

Even when two platforms see the same engine output, they can disagree about whether the brand was cited. Edge cases that matter: a URL is cited but the brand's name doesn't appear in the answer text — counts as a citation? A competitor's site mentions the brand by name and gets cited — counts as a brand citation? A redirect chain ends at the brand's domain — counts? The brand's social profile (LinkedIn, Twitter) is cited — counts as a brand citation or a third-party one?

Platforms make different choices, often without documenting them. Two reasonable engineers can land on different policies, and the resulting numbers diverge.

The question that surfaces it: what specifically counts as a citation event in the platform's definition?

The five questions to ask any AEO platform's measurement layer

Put the four sources of disagreement together with a fifth question about reproducibility, and you get the five questions worth asking any vendor:

• Prompt set: how many prompts, where did they come from, refresh cadence, and were they tuned for my category?
• Cadence: how often is each prompt re-queried, and how is intra-day engine variance handled in the reported number?
• Non-determinism: how does the platform handle same-prompt re-run variance — single sample, multiple samples averaged, or something else?
• Citation definition: what specifically counts as a citation event, including the edge cases (URL-only cites, third-party mentions, redirects, social profiles)?
• Reproducibility: can the platform reproduce a single reported number — same prompt, same time window, same engine, same parser — and show me the underlying data?

How SolCrys answers each — with uncertainty disclosed

Here are our honest answers.

What this doesn't fix

Five answered questions don't resolve every measurement disagreement. Different platforms making different reasonable choices on the four dimensions will still produce different numbers, and neither has to be wrong. What changes after this conversation is that you know which choices each platform made and can decide which set of choices fits your team's question.

Concretely: if your CMO's quarterly question is "are we trending up across the engines that matter for us," a 7/30-day rolling sample on a stable curated prompt set (our choice) is probably the right design. If your question is "did we move on this specific prompt today," a high-cadence per-prompt sample with multiple re-runs (a different design) might fit better. The platforms are answering different questions. That's a feature.

FAQ

Related guides