How Prescy measures AI recommendations

What Prescy checks

Monitor repeatedly sends 60 core buyer questions to two engines each week: up to 5 samples on ChatGPT and 5 on Perplexity. Every question gets three ChatGPT samples; the remaining samples are added only where those four disagree on whether the business was named, so the extra cost goes to the questions that are unstable. Prescy stores every available answer so you can inspect the evidence behind the result.

It is ChatGPT and Perplexity specifically because those are the only two engines you can query reliably and get real, machine-readable citations back from. The ranked source queue — the pages deciding your category, and what the rest of the product is built on — depends on those citations. An engine that can only be scraped unreliably, or that returns an answer with no traceable sources, would add to the coverage count without making the source queue any better.

Each weekly Monitor benchmark aims for 600 primary observations: 60 journeys × (5 ChatGPT + 5 Perplexity). We sample more where the answers vary: ChatGPT can answer the same question differently from run to run, so its extra samples go to the questions where the first runs disagree. Perplexity is retrieval-driven and more stable, so its repeated answers tend to corroborate each other rather than add independent information. Google AI Overviews is not measured — there is no reliable, terms-safe way to access a surface that changes without notice and triggers inconsistently, so a number for it would not be trustworthy. Track Overview impressions in Search Console instead.

The searches being measured

During setup, Prescy suggests buyer-style journeys from your website and asks you to confirm what you sell, who you want to win and your likely competitors. Journey editing remains available after activation as an advanced control.

Mentioned, considered or recommended

A result counts as a mention when the available answer contains your business name, a saved business alias, your full domain or an unambiguous domain label. Matching is case-insensitive and uses word boundaries so a short business name is not counted inside an unrelated word.

The paid headline rate pools the repeated ChatGPT and Perplexity observations. Each successful answer on either engine counts as one tested buying conversation, so questions with unstable answers contribute more observations than steady ones. Plain mentions never increase it.

Samples below 50 tested conversations are shown as raw counts because a percentage would imply false precision. At 50 or more, Prescy shows an approximate percentage alongside the full sample size. This threshold is a conservative presentation rule, not a guarantee of statistical significance.

Per-engine visibility

Each engine's visibility percentage is the number of applicable answers that mention your business divided by the number of applicable answers successfully checked for that engine.

Example: if your business appears in 7 of 20 available ChatGPT answers, ChatGPT visibility is 35%.

Why there is no 0–100 score

Prescy does not compress the result into a proprietary 0–100 score. Composite scores hide the denominator, are calculated differently by every vendor and can make a thin sample look more certain than it is.

The numerator, denominator, buyer question and original answer remain available. Recommendation outcomes, plain mentions, cited sources and competitor appearances are kept separate so a reader can inspect what actually moved.

Missing, delayed and unavailable results

• A request that could not run is marked unavailable and excluded from the reported denominator.
• If ChatGPT or Perplexity returned no usable answer for a search, that engine's run for that search is not treated as a measured absence.
• If only some requests succeed, each engine's percentage uses its own successfully available answers as the denominator.
• Perplexity is queried in a follow-up job just after ChatGPT, so a scan can briefly show the ChatGPT-only count before the pooled number settles.

The dashboard displays how many searches were included per engine, so two scans with different coverage can be interpreted correctly.

Competitor measurement

Competitor results combine appearances in your tracked buyer searches with bounded, paid-plan comparison searches. Those comparison searches include head-to-head and alternatives questions. A competitor being named in the question itself does not count as visibility; Prescy looks for its appearance in the answer.

Competitor queries are stored separately and never change your own business's recommendation count. Competitors are matched by the name you provide, so aliases that have not been entered may be missed.

How recommendations are produced

Recommendations are matched to a failure stage, not a generic content rewrite. Prescy first checks whether live retrieval crawlers can reach the site (robots.txt and CDN-level overrides). It then classifies each lost buyer question as: no search fired, searched but not retrieved, retrieved but not cited, or cited unfavorably. The action has to fit that stage.

Every action carries an evidence grade (strong, mixed, weak, vendor-only, disproven, or hygiene) and a predicted direction. Schema markup and llms.txt are labeled hygiene — they are not sold as visibility levers. Prescy does not recommend hidden text, fabricated statistics or reviews, memory-poisoning tricks, or a promised AI rank or traffic number. For the fuller argument, see what AI search actually is.

A reasoning model still writes the per-question diagnosis (winner, excerpt, why). Evidence quotations are checked against the actual source text. The action itself is assigned by the stage classifier so the default is never “rewrite your content.”

How recommendation outcomes are measured

When you mark an action implemented, Prescy keeps the full scan that produced it as an immutable baseline. Later results only count when a full scan contains the exact same tracked customer search. The first qualifying rescan is labeled an early signal; two consecutive results pointing in the same direction are labeled stronger evidence.

Improved timing does not prove the action caused the movement. When a held-out control prompt is available, Prescy reports whether that untreated question moved in the same window — including when the treated and control results are statistically indistinguishable from a tie or from platform drift.

How the AI Search Report expands

The free AI Search Report starts before signup. Prescy uses the business website to suggest eight representative buyer questions, checks each question once in ChatGPT, shows whether the business appeared and keeps the original answers available as evidence.

Saving the report with a free account preserves the diagnosis and expands the same journey. Paid monitoring adds a repeated ChatGPT and Perplexity benchmark, deeper competitor reasons, more actions and history, so the user can see what changed.

Limits and interpretation

• AI answers are probabilistic and can vary between otherwise identical requests.
• The counts measure the selected searches and the available ChatGPT and Perplexity answers—not every possible customer question or AI product.
• A higher mention or recommendation count does not prove preference, sentiment, accuracy, traffic or revenue.
• Provider model changes can create movement even when your website has not changed.
• With only two engines, agreement between ChatGPT and Perplexity is a corroboration check, not cross-engine consensus. A source or answer that comes through on both is a stronger signal than one that only repeats on a single engine.
• 5 ChatGPT and 5 Perplexity samples per core question are enough to expose variation, not to estimate a stable per-question probability. Weekly movements are signals to inspect; trends over several scans are more meaningful.