AI Visibility Tracking From GSC Queries: A Practical System

Google Search Console can report clicks, impressions, CTR, and average position for the queries that surfaced your site, but it does not tell us whether ChatGPT or Perplexity named our brand in the answer. This AI visibility tracking workflow turns the GSC queries we already own into repeatable prompts, giving us a practical way to measure mentions, citations, competitor gaps, and share of answer by engine.

A Reddit discussion about finding queries a site is “ranking on” in AI captures the common constraint: GSC is accessible, while enterprise AI-search platforms can be costly. We should not relabel GSC impressions as AI rankings. Instead, we use GSC to prioritize real audience demand, then test the resulting questions under documented, repeatable conditions.

GSC is source data, not proof of AI visibility

Google Search Console’s Search results Performance report includes clicks, impressions, CTR, average position, queries, pages, countries, and devices. Those are useful SEO fields, but an impression means a result from our site appeared in Google Search; it does not establish that an AI-generated answer mentioned our company, cited our URL, or preferred us to another provider.

Google’s guidance for AI features also says that traffic from AI features is included within the overall Search Console Web search reporting rather than presented as a separate, prompt-by-prompt brand-mention report. In practical terms, we can use verified GSC fields to understand search exposure, but we must observe and log answer content separately if we want to know which brands appeared.

These are different claims and should remain different in a report:

• “Our URL received a Google Search impression.”
• “Our page gained clicks from a query associated with buyer demand.”
• “ChatGPT named our brand in a response.”
• “Perplexity displayed a citation to a page on our domain.”
• “Our brand appeared first in an explicitly ordered vendor list.”

The first two are Search Console measurements. The final three are prompt-level observations, collected from an AI-answer surface or API response under our test protocol. For a fuller distinction between output visibility and source links, see our guide to AI citations vs. AI visibility tracking.

Define the metrics before starting AI visibility tracking

A durable dashboard separates conventional Google signals from answer-level signals. Combining impressions, rankings, mentions, and citations into one unlabeled chart makes it difficult to tell whether a change came from organic search exposure, a model response, or a changed prompt set.

Traditional rankings and impressions

Average position is Search Console’s calculated position metric for Google results, while impressions indicate that a result was shown. For example, a project-management platform might have an average position near 4 for “best project management software for agencies” and receive 8,000 impressions in a reporting period. That does not prove the brand appears in a response to the equivalent ChatGPT prompt.

Use these fields to select and prioritize questions, connect queries to landing pages, and investigate SEO changes. Do not use them as a proxy AI visibility score.

Mentions, citations, coverage, and share

For each prompt-engine-run, we recommend recording these distinct outcomes:

1. Brand mention: our normalized brand name appears at least once in the answer.
2. Brand-domain citation: the engine visibly displays a source URL or domain we own.
3. Prompt coverage: the percentage of eligible observations where our brand is mentioned.
4. Citation coverage: the percentage of eligible observations with a visible brand-domain citation.
5. Share of answer: our entity presence relative to all eligible named vendor entities in the answer.

Ahrefs’ AI Visibility Metrics documentation distinguishes AI mentions, citations, impressions, and share-of-voice-style measurements. That distinction is useful even when we do not use Ahrefs: a visible citation is not the same thing as a mention, and neither is the same thing as conventional search visibility.

Export GSC queries that reflect buyer demand

Start with a 3- or 6-month GSC export from the Search results Performance report. Repeat the same export on a regular schedule, such as monthly, and preserve the date range so later comparisons remain interpretable.

Include these columns in the working file:

Do not automatically take the top 100 queries by impressions. Branded navigation, misspellings, and near-duplicate variants can dominate a raw export without teaching us much about early-stage buyer discovery.

Instead, label each query by intent. A 50-prompt starter set might contain 15 commercial-investigation terms, 10 comparisons, 15 problem-led informational terms, 5 category-definition terms, and 5 local-service terms. The exact mix varies by business, but writing the selection rule down prevents cherry-picking later.

Examples include:

• Commercial investigation: “best payroll software for small business”
• Comparison: “Rippling vs Gusto”
• Problem-led informational: “how to reduce employee onboarding time”
• Category definition: “what is payroll automation”
• Local service: “emergency electrician in Austin”

Commercial and comparison questions are sensible priorities because users explicitly frame them as vendor-selection requests. We should not assume every assistant will provide recommendations; the test is designed to observe what each engine actually returns. If GSC data changes abruptly, check reporting context before changing the prompt set; our analysis of the September 29 Search Console data gap explains why an apparent drop may not equal lost demand.

Turn each GSC query into a stable prompt

A raw search query is often too fragmentary for consistent AI testing. “CRM agency” and “best ai tool seo” can be interpreted several ways, so testing the fragment one week and a detailed version the next produces a trend line we cannot defend.

We preserve the underlying need while converting selected queries into clear, stable questions.

Apply a written normalization rule

Add only qualifiers necessary to make the request reproducible: location, language, target customer, use case, or a budget band where the original query supports it. Do not insert a qualifier merely because it favors our brand.

Store the original query and normalized prompt together. For local businesses, geographic context is essential: “best family lawyer” and “best family lawyer in Phoenix, Arizona” are separate prompts, not variations to average together. Our guide to testing ChatGPT local search citations for top list pages covers the additional care needed for local list-style prompts.

Keep meaningful variants separate

Do not merge “best email marketing platforms,” “best email marketing platform for a 10-person ecommerce team,” and “MailerLite vs Klaviyo for Shopify stores.” They represent category discovery, segment-specific selection, and direct comparison. An absence on a broad prompt may coexist with strong visibility in a high-value segment, which is a business finding rather than a measurement error.

Test AI engines with a documented protocol

A single browser test is useful research, not longitudinal measurement. Responses can vary with wording, model or product updates, location, language, search access, available sources, and prior chat context.

Semrush’s published AI Visibility Index methodology describes AI visibility as a modeled measurement based on its own datasets and methodology. We should treat any platform score, including a score we calculate ourselves, as a defined snapshot of a particular prompt set and test environment—not as a universal ranking.

For every observation, capture six controls:

1. Engine and surface: ChatGPT, Claude, Gemini, Perplexity, Grok, or a separately observed Google Search feature.
2. Model or version identifier: record it when the product exposes one.
3. Market context: country, city for local prompts, and language.
4. Exact prompt: including qualifiers and punctuation.
5. Run date and time: for example, 2026-10-11 14:00 UTC.
6. Session rule: fresh session, no account history where controllable, and no follow-up prompts.

For a 50-prompt set across ChatGPT, Gemini, and Perplexity, one scheduled cycle creates 150 observations. Weekly testing is a reasonable initial cadence; daily testing may be justified for a smaller set of critical prompts, but it creates more noise to review.

We also avoid treating a Gemini API response as interchangeable with a Google AI Overview. They are different product surfaces and may use different systems. Google-specific answer observations should be labeled as manual or tool-assisted observations, while GSC fields should retain their Google-provided names and definitions.

Record answer evidence in a compact template

One row per prompt, engine, and run date is enough to begin. A spreadsheet, CSV, or local database can support a useful system before a team needs a larger reporting stack.

Count a brand mention once per entity per answer, even if the name appears five times. Use an approved alias list—for example, a legal company name, common product name, and known abbreviation—so “Acme,” “Acme Inc.,” and “Acme CRM” are not inconsistently coded.

Capture citations literally. If an engine displays a URL, source card, or domain, store what was visible. Do not infer a citation because the answer resembles our page. This answer-level discipline is one reason query fan-out analysis and AI visibility tracking are separate tasks: related-query discovery can inform research, but it does not prove final-answer inclusion.

Calculate an auditable AI visibility score

There is no industry-standard AI visibility score. Semrush, Ahrefs, and other vendors use their own data, prompt universes, and methodology choices, so scores from different systems are not directly exchangeable.

For our own stable prompt set, a transparent weighted score is more useful than an opaque composite. Apply these objective codes to each eligible prompt-engine-run:

• 0 points: our brand is absent.
• 1 point: our brand is named at least once.
• 2 points: our brand is named and a brand-owned URL or domain is visibly cited.
• 3 points: our brand is named in position 1 of an explicitly ordered vendor list and a brand-owned URL or domain is visibly cited.

Do not award 3 points to narrative prose such as “we recommend Acme,” because preference language can be ambiguous. Record that wording in notes or create a separately defined recommendation flag if it matters to the business.

AI visibility score = total points earned / total possible points × 100

If 40 prompts are tested across 3 engines, there are 120 observations and 360 possible points. A total of 108 points is a 30.0 score. Keep the prompt list, entity rules, engines, market settings, and cadence stable; if we add 20 prompts in November, mark a series break instead of presenting an artificial month-over-month comparison.

Define share of answer without hiding missing data

Share of answer needs an explicit denominator. We use an entity-level method: within each eligible answer, count each named vendor entity once, then aggregate those counts across the reporting period.

The denominator includes our brand, each tracked competitor, and an Other named vendor bucket for entities outside the tracked set. Repeated mentions do not add weight. Generic, unnamed recommendations such as “a local specialist” are excluded from the entity denominator because no brand can be assigned, but we flag the answer as unnamed recommendation present and report that rate separately.

For example, across 20 eligible answers, our brand appears in 6, Competitor A in 8, Competitor B in 4, Competitor C in 2, and untracked vendors in 5 answers. The denominator is 25 named entity appearances, so our share of answer is 6 ÷ 25 = 24%. If 4 of the 20 answers make an unnamed recommendation, report a 20% unnamed-recommendation rate alongside share rather than silently excluding that context.

This prevents a narrow competitor list from inflating our share. It also distinguishes entity share from prompt coverage: we might have 24% entity share while appearing in only 15% of tested answers.

Use competitor gaps to decide what to investigate

The useful output is not simply “we had 42 mentions.” It is a ranked list of buyer questions where competitors repeatedly appear and we do not.

Create a gap view with these filters:

• Our brand absent.
• At least one named competitor or Other named vendor present.
• Commercial or high-intent informational intent.
• The same competitor appears in two or more engines or test dates.
• A visible source, cited page, review, directory, comparison, or editorial result is recorded where available.

For example, an accounting-software company may be absent when ChatGPT and Perplexity answer “best accounting software for freelancers,” while a competitor appears on multiple test dates. We then inspect the observed associations: freelancer pricing, invoicing workflow coverage, integrations, editorial comparisons, or third-party reviews. We do not assume that adding the exact phrase to a landing page will change the answer.

Semrush’s AI Visibility solution describes tracking brand presence across AI platforms, while Ahrefs documents separate AI visibility metrics. Those products can suit teams needing broad vendor datasets or managed workflows. Our local-first approach starts with the questions our audience already asks and preserves the customer’s API key, prompt list, and raw response evidence. That query-level control is especially practical for an agency managing 20 accounts or a brand beginning with 50 decisive questions.

Report changes without confusing AI and Google signals

A monthly dashboard should keep Google Search data and AI-answer observations side by side, not blend them into one score. For a 50-prompt, 3-engine program, report the 150 scheduled observations, successful runs, and any failed or unavailable runs before calculating rates.

A practical dashboard has six blocks:

• GSC clicks, impressions, CTR, and average position as SEO context.
• Observation completion rate by engine.
• Brand mention coverage by engine.
• Brand-domain citation coverage by engine.
• Share of answer, including the Other named vendor bucket.
• Top 10 competitor-gap prompts and their visible evidence.

Suppose 45 of 150 observations mention our brand, and 12 of those show a visible citation to our domain. Mention coverage is 30.0%; citation coverage is 8.0% of all observations and 26.7% of brand-mention observations. Those labels make the calculation reviewable.

AI Visibility Tracker supports this workflow locally across ChatGPT, Claude, Gemini, Perplexity, and Grok using the customer’s own API key. It does not turn conventional rankings into AI mentions; it helps us retain the exact prompt, engine, date, answer evidence, and competitors needed to assess answer-level visibility over time.

FAQ

How can I find which queries generate AI mentions or citations for my brand?

Export relevant Google Search Console queries, classify them by intent, and convert the priority queries into stable prompts. Test each prompt under the same documented conditions across the engines that matter to your buyers. Log the exact answer outcome: brand mention, visible brand-domain citation, named competitors, named entities outside your tracked list, and answer format. GSC supplies candidate demand; prompt testing verifies answer presence.

Can Google Search Console show the queries where a brand appears in AI-generated answers?

No. Search Console can show query, page, click, impression, CTR, position, country, and device data for Google Search reporting, but it is not a complete transcript of brands named in AI-generated answers. Keep GSC metrics as Google Search context. Use a separate observation log to record whether a specific tested answer named your brand or visibly cited your domain.

How do I turn GSC queries into prompts for tracking AI visibility?

Retain the original need but rewrite fragments into stable natural-language questions. Add only necessary qualifiers such as country, customer type, use case, language, or location. For example, convert “best crm agencies” to “What are the best CRM platforms for marketing agencies in the United States?” Store both versions, and do not merge broad category prompts with detailed segment or comparison prompts.

How is an AI visibility score calculated?

There is no universal score, so use rules your team can reproduce. Our example assigns 0 for absent, 1 for a brand mention, 2 for a mention plus visible brand-domain citation, and 3 only when the brand is first in an explicitly ordered vendor list with a visible citation. Divide earned points by possible points, then preserve the same prompts, engines, and coding rules for comparisons.

What is the difference between rankings, impressions, AI mentions, citations, and share of answer?

Traditional rankings and impressions are Google Search metrics: position estimates where a result appeared, while an impression shows it was displayed. An AI mention means the response named the brand. A citation means a visible source linked to or identified the brand’s domain. Share of answer is the brand’s share of named vendor entities in tracked answers, including an Other bucket for untracked vendors.