Quick answer

A reliable AI visibility dashboard must answer four separate questions: is the brand mentioned? is its domain cited? does the user visit the site? does that influence contribute to a business outcome? A score that blends these steps hides the denominators, the volatility, and the real value.

Start with a versioned prompt panel, stratified by intent, and run several times on each platform. Measure mention rate, citation rate, share of voice, and stability. Then add proprietary data: Search Console or Bing AI reports where available, Analytics referrals, CRM conversions, and brand searches. Keep platforms separate; their surfaces, reports, and methods aren't comparable unit for unit.

Key takeaways

  • A citation count with no number of prompts, runs, and platforms isn't interpretable.
  • A mention names the brand; a citation attributes a source; a visit and a conversion are later events.
  • Search Console and Bing Webmaster Tools provide different proprietary metrics; don't add them together.
  • Share of voice depends on the chosen panel: it must always show market, language, intent, and version.
  • A useful measurement leads to a decision: reinforce evidence, fix access, improve a page, or reallocate effort.

The seven metrics to keep separate

Metric Formula / source Possible decision Main pitfall
Mention rate runs with brand / valid runs Verify the entity's presence A mention can be negative
Citation rate runs with a domain URL / valid runs Identify source pages A citation can generate zero clicks
Prompt coverage prompts with at least one presence / unique prompts Find absent intents Hides volatility between runs
Share of voice brand's presences / panel brands' presences Compare at a constant scope Depends entirely on the panel and the counting rule
Volatility prompts changing status / repeated prompts Determine the number of repeats Can be confused with a model change
AI referral sessions identified by source/medium Analyze on-site behavior Underestimates influence with no click or misattributed
Business outcome leads, sales, or revenue per CRM rule Prioritize profitable tasks Non-causal attribution and multi-touch journeys

Don't add these metrics together. A brand can hold 20% share of voice, 5% citation rate, and 0.2% referral traffic: each answers a different question.

What platforms state they measure

Google Search Console

On 3 June 2026, Google announced Search Generative AI reports dedicated to AI Overviews and AI Mode impressions, with pages, countries, devices, and trends over time. The rollout initially covered a subset of sites (Google's announcement). The help documentation notes familiar limits: restricted availability, thresholds, canonicalization, aggregation, and a row cap. An impression indicates visibility per the report's own definition; it doesn't give sentiment, a recommendation, a click on a source, or a conversion.

Bing Webmaster Tools

In February 2026, Bing launched an AI Performance report in public preview: total citations, average cited pages, sampled grounding phrases, and URL trends on Copilot, Bing AI summaries, and certain partners. Bing warns that these figures indicate neither position, nor authority, nor rank (Bing Webmaster Blog). It's a different scope; a Bing citation can't be added to a Google impression as if they shared a unit.

Multi-platform trackers

A tracker generally runs a prompt panel and codes the answers. Its value depends on verifiable parameters: exact text, source of the prompts, country/language, account, frequency, repeats, error handling, mention/citation rules, response archiving, model changes, and exports. If these parameters are opaque, the proprietary score can't be audited.

Semrush, in its 2026 methodology, distinguishes mention from citation and analyzes 126 million deduplicated US prompts across ChatGPT, Google AI Mode, AI Overviews, and Gemini. The index remains proprietary and doesn't publish every execution or uncertainty parameter; its good practice is to keep platforms separate (Semrush AI Visibility Index methodology).

Building a measurement system in six steps

1. Write the decision question

Example: "Are the comparison pages we publish earning non-branded citations and demo requests?" This wording avoids a dashboard with no action.

2. Define the panel

Split prompts between discovery, problem, comparison, and decision. Mark branded/non-branded, country, language, audience, and business value. Keep a fixed core panel; put new questions in an exploratory cohort.

3. Fix the protocol

Choose platforms, repeats, cadence, error rule, and data window before the test. Note the model version when visible. The protocole-visibilite-ia.csv file provides the raw structure.

4. Code and check

Automate initial detection, then review a sample. Homonyms, subdomains, navigation links, and negative citations produce false positives. Measure agreement between two reviewers on 10% of answers if the stakes are high.

5. Reconcile with proprietary data

Connect cited URLs to Search Console, Analytics, and CRM without forcing equality. Google notes that Search Console observes performance before the visit, while Analytics observes the site; canonical, timezone, and attribution explain legitimate gaps (Google Search Central).

6. Version and decide

Compare periods with an identical protocol. A change in model, panel, or method creates a series break. Every report ends with three decisions, their owners, and their dates, not a decorative score.

Example of an honest scorecard

A company tests 100 prompts, five times, on three platforms: 1,500 expected runs. Thirty errors are excluded per a pre-established rule; the valid denominator is 1,470. The brand appears in 294 answers (20%), the domain in 88 (6%), and 34 of 100 unique prompts get at least one citation (34% coverage). The status changes between runs for 27 prompts (27% volatility).

Result Value Reading Decision
Mention 20% The entity is known within the panel Examine context and prompts with no citation
Citation 6% The site sometimes serves as a source Identify the 88 URLs and the reused evidence
Coverage 34% A third of intents get at least one citation Map the 66 absent intents
Volatility 27% A single run would be fragile Keep five runs and publish a range
AI sessions 42 Small, identifiable volume Compare engagement, without concluding incrementality
Demos 3 Possible commercial signal Verify CRM, journey, and cohorts before attribution

A composite score could turn these values into 72/100, but the reader would no longer know whether the improvement came from a branded mention or revenue. The scorecard keeps causalities separate.

What the data proves and doesn't prove

Seer analyzed 53 brands, 5.47 million complete queries, and 2.43 billion organic impressions from January 2025 to February 2026. In this cohort, brands cited on SERPs with AIO earned 120% more organic clicks per impression than uncited brands, but 38% fewer than queries with no AIO (Seer Interactive). The most recent AIO/citation status was applied to historical months, the panel comes from clients, and the selection isn't randomized. This is a useful association for segmenting, not proof that citation causes the click.

Adobe observed over a trillion US retail visits from January to March 2026 and reported that AI-referred visits had grown 393% year over year for the quarter; in March, they converted 42% better than non-AI traffic (Adobe). The absolute level isn't published, the base can be small, the scope is US retail instrumented by Adobe, and "non-AI" groups heterogeneous channels. Don't transpose this ratio to a French SaaS.

Common mistakes and safeguards

  • Publishing a raw citation count. Add prompts, runs, errors, platforms, and period.
  • Mixing up brand and domain. A mention with no URL and a citation are different events.
  • Comparing two trackers on their scores. Compare panel, repeats, and counting rules first.
  • Changing the panel every month. Keep a stable core and a separate exploration cohort.
  • Treating an absence as a technical bug. It can come from the query, selection, evidence, or access; diagnose in that order.
  • Proving ROI with last-click only. Show referrals, assisted conversions, and brand searches as distinct angles.
  • Storing answers with no governance. Limit personal data and document retention, access, and deletion.

How SEOryon fits in

SEOryon tracks visibility in ChatGPT, Perplexity, Gemini, and Claude, connects that observation to content recommendations, and can use read-only Search Console/Analytics. The free score audits a URL on 27 deterministic signals. For credible reporting, keep the raw values, the panel, and the protocol version; use the platform to automate tasks, not to hide the assumptions.

Measurable exercise

Build a monthly scorecard with 50 prompts, three platforms, and four runs. Calculate the seven metrics in the table, segment by intent, and write three decisions with an owner. Success: a third party can recompute the rates from the CSV, and every change states whether the panel or the model changed.

Where to go next

FAQ

Is there an official AI visibility score?

No. Google and Bing publish certain proprietary metrics; multi-platform indices are vendor constructs. Demand their method and the raw values.

How many prompts should you track?

Enough to cover the intents that change a decision, not to hit a marketing number. Fifty stratified, repeated prompts often teach more than a thousand opaque ones.

Should you count several citations in the same answer?

Decide before the test. For the citation rate, a run can be worth 0/1. For URL share of voice, you can count occurrences; publish that rule then.

Does Search Console show AI Overview clicks?

The generative report announced in June 2026 centers on impressions and is available on a subset of sites. Check your property's documentation; don't infer a conversion from an impression.

How do you measure influence with no click?

Use separate proxies: brand searches, "how did you hear about us?" surveys, geographic or time-based tests, and business data. Present them as clues, not certain attribution.

References

Method note

Method: documentation and studies verified 16 July 2026, translated and edited 22 July 2026. Preview reports, models, and tracking rules can change; version any series break.