Methodology

How SearchIntel measures AI visibility

We sample six AI platforms, label every brand-present answer against a published rubric, and grade each evidence panel against a five-input confidence model. This page lists every input, weight, and caveat the product uses so the numbers you see can be audited.

Last updated 2026-05-26. Methodology versioned alongside the product release.

Six AI platforms we sample

Every monitor queries six platforms per keyword. Five LLMs receive a natural decision prompt (e.g. "What are the best X for Y?") and Google's AI surfaces are captured via SerpAPI's official endpoints.

AIO Google AI Overviews

SerpAPI google engine. Captures the AI Overview block + linked source URLs when Google returns one.

AI Mode Google AI Mode

Launched 19-20 May 2026 at Google I/O. SerpAPI google_ai_mode engine. Continuous-refresh layer distinct from AIO.

GPT ChatGPT

OpenAI gpt-5-chat-latest, the rolling alias that always points to the current ChatGPT consumer default (GPT-5.5 Instant as of May 2026). When OpenAI ships a new default, our sampling follows automatically, so we always measure the model the ~900M weekly ChatGPT users actually see. 5 runs per keyword for stability scoring.

Claude Anthropic Claude

Anthropic claude-haiku-4-5. 5 runs per keyword. Anti-recommendation hedge tuning matches the rubric below.

Gemini Google Gemini

Google gemini-3.5-flash (new default since Google I/O May 2026, with 2.5-flash as fallback). 5 runs per keyword.

Perplexity Perplexity

Perplexity Sonar online. Captures native citation URLs alongside the answer text.

Per-keyword market context (region gl + language hl) is applied uniformly across all six platforms. Six market presets ship: UK, US, DE, FR, ES, NL.

Evidence confidence grade, five inputs

Every Workflow Summary surfaces an Evidence Confidence letter grade. The grade is a weighted composite of five inputs. Each input scores 0-100; the composite is graded A → F. Weights are locked in the versioned confidence model and audited per release.

1. Run cadence

25%

Daily over 28+ days → 95. Weekly cadence → 78. Sparse / 1-2 runs → 40-60. Encourages cadence so trend signal is meaningful.

2. Platform coverage

20%

All six platforms returning answers → 100. Platforms with zero successful captures pull the score down proportionally.

3. Prompt sample size

20%

≥200 prompts → 100. Diminishing returns below that; <25 prompts triggers the Sample-limited cap that holds the display grade at B+ regardless of composite.

4. Evidence-linked claims

25%

Share of AI answers that linked to a source URL we could capture and trace. AIO + AI Mode + Perplexity contribute the most; LLM mentions without links count partially.

5. Volatility (lower is better)

10%

Stability of brand-present signal across the 5-runs-per-keyword sampling. High volatility (e.g. visible on 1 of 5 runs) reduces this score; perfect stability (5 of 5) maxes it.

Composite → letter grade

A: ≥92 A-: ≥87 B+: ≥82 B: ≥76 B-: ≥70 C+: ≥64 C: ≥58 C-: ≥50 D: ≥40 F: <40

Sample-limited cap

When the Prompt Sample input is below grade D (under 25 prompts), the displayed letter is capped at B+ regardless of the raw composite. The raw composite is preserved on the methodology drawer. This keeps a brand-new monitor with three keywords from showing an A.

Recommendation classifier, four-label rubric

The recommendation classifier reads one AI answer at a time and assigns one of four labels. The full rubric, phrase triggers, anti-triggers, and worked examples, is published as the four-label summary below.

Recommended

Brand is affirmatively endorsed as a top option. Phrase triggers: "I recommend", "go with", "the best", "top pick", positioned first in a non-negative ranked list, or "excels in / ideal for" contrast.

Mentioned

Brand appears in the answer but without endorsement. Listed alongside other options, hedged with "but / however / although", or factually described without a recommendation frame.

Negative

AI warns against the brand. Phrase triggers: "avoid", "don't use", "instead of [brand]", "[brand] is not recommended", or a "watch out for" clause specifically about the brand.

Absent

Brand does not appear in the answer text at all. Calculated as the residual: total answers − (recommended + mentioned + negative).

Calibration evidence

Current calibration set: 82 hand-labelled answers across Monzo + SearchIntel (May 2026). Classifier-vs-human agreement: 81/82 = 98.8%. Negative-recall floor: 100%. Per-paying-client vertical validation is staged separately; until that fires, any new monitor surfaces the rubric counts as a calibration pilot with the caveat copy locked on the drawer.

The single disagreement is an unknown-brand acknowledgement edge case the phrase-rules classifier reads as mentioned where the human labeller chose recommended. Tracked as a known-miss for the next classifier version.

Anti-triggers (what we don't count as recommended)

Visibility score formula

The headline visibility score on every monitor + the free /check is a single percentage. Formula: 

score = round(brand_visible_count / (keywords × platforms) × 100)

On AI-Mode-aware monitors platforms = 6. On older runs that pre-date the May 2026 AI Mode launch, platforms = 5. The score is always relative to the platform count of THAT run, we never retroactively inflate older scores.

Example: 20 keywords × 6 platforms = 120 possible brand-visible slots. If the brand was visible in 36 of those, the score is round(36 / 120 × 100) = 30.

Sampling + run cadence

Collection + output integrity checks

Every scan output passes three integrity checks before any client deliverable. These are scripted, not eyeballed.

Every regenerated output must re-derive from the current scan data, no phrase or number survives a re-run.

Caveats + what we don't claim

Not a market-wide claim. The visibility score reflects the SAMPLED query set, not the entire universe of AI queries for the vertical. A 35% score against 20 prompts is one signal, directionally true, not a market census.

Not measured ad-platform traffic. AI demand estimates are modelled from search demand correlation + query type, not from a panel of real AI users. We surface this caveat on every demand-estimate panel.

No production-grade recommendation rate per paying-client until vertical-specific calibration. The 82-row cross-client validation gives us a directional rubric. A new paying client gets the recommendation rate framed as a calibration pilot until that client's own labelled validation run is committed.

Brand-presence detection has known noise. Generic English words occasionally surface as brands in low-context LLM responses (e.g. "Extra", "Max", "Perks" in financial verticals). We surface this as a transparency note and recommend client review of detected competitors.

Methodology change log