Methodology

How MentionFlow measures AI search

Visibility data is only useful if you can defend it. This page documents exactly how we sample each engine, how every metric is computed, and where the limits are. When the methodology changes, we version it and annotate your charts.

• Every tracked prompt runs on every selected engine once per day (plan-dependent cadence), per configured region with its local language and Google location targeting.
• We sample consumer surfaces wherever they diverge from developer APIs — our published bake-off found official-API ChatGPT answers agreed with the real surface only 42–67% of the time on brand presence.
• Raw answers are stored verbatim and re-scored whenever our extraction improves (it has happened twice already; both re-runs are in the changelog).
• All metrics report on rolling windows — never single runs. Windows with fewer than 5 samples are flagged low-confidence.

• Two passes: a deterministic alias matcher decides presence, then an LLM pass ranks every brand in the answer (tracked and untracked), assigns sentiment, and quotes the verbatim evidence behind it.
• mention position means rank among all brands in the answer — what a buyer actually sees.
• Measured on a hand-labeled golden set: presence F1 100, sentiment agreement 84% and climbing as the set grows.
• Sponsored content policy: ads that engines display inside answers are captured, excluded from every metric, and shown separately on each answer's receipt.

The eight above are the search-grounded engines — they retrieve the live web before answering, and they are the only ones every blended headline metric is computed over. Three more answer purely from what the model absorbed in training:

Their APIs have no web-search step, so their citations are honestly empty — not missing data, an absence of anything to cite. That is also why they are excluded from the blend: a memory-only mention measures something different from a retrieved one, and averaging the two would give the headline number two meanings. Per-engine views show them in full, and filtering to one of them shows that engine unblended.

Nobody outside the AI labs knows how often a specific prompt is asked. OpenAI, Google and Anthropic don't sell their logs; vendors claiming "real prompt volumes" are projecting from small browser-extension panels. So we don't pretend. Our estimate is a model, and here is exactly how it works:

• Start from what assistants actually did. When an AI engine answers a tracked prompt, it usually runs web searches first. We record those exact queries on every collected answer — observed behavior, never simulated.
• Measure real Google demand for them. Those queries (plus the prompt's own head keyword) are looked up against Google Keyword Planner volumes — real search demand, not an index. The estimate is anchored on the strongest matched query (the maximum, not a sum: most engine-rewritten queries have zero measurable volume, and summing would reward that noise).
• Scale to AI usage — as a band. Published research places search-like AI asks at roughly 4–12% of comparable Google volume. We show that whole range (e.g. "1.2K–3.5K/mo"), never a false-precision integer.
• Calibrate against your own Search Console. When your site is connected, we match those same queries against what Google itself reports for your site. A match is real demand evidence from your market: the prompt is labeled measuredand we show your site's actual impressions on that query over the last 28 days — a real number, next to the modeled band, never blended into it. GSC impressions are how often your site appeared on Google for that query, not how often the query is asked globally, so the band stays a band.
• Label the confidence. measured = a query behind this prompt appears in your own Search Console data. anchored = a search an assistant actually ran for this prompt matched real Google demand. modeled = only the prompt's head keyword matched. "—" = no measurable demand signal yet — which is information, not zero.

This is a modeling choice, not a measured relationship: we don't claim a validated link between Google search volume and how often a prompt is asked inside an AI tool — and neither can anyone else. The only exception is labeled as such: measured impressions come from your own Search Console and count your site's Google visibility, nothing more. We'd rather show you an honest band than an invented count.

• LLM answers are non-deterministic; day-to-day movement inside a window is expected noise. Judge trends, not ticks. Measured on our own tracked brands (1,034 prompt×engine consecutive-day pairs, July 2026): 92.6% kept the same brand presence between consecutive days. Per-engine flip rates ranged from ~4% (Perplexity) to ~11% (AI Overviews). We re-run this number as the dataset grows.
• We sample logged-out, so personalization and memory are excluded. We consider this a feature for measurement, and we say so out loud.
• Engines ship model updates without notice. We record the model version on every run and annotate charts when versions shift.
• US-English sampling by default; additional locales are sampled separately, never blended.