AI visibility question · direct answer

How Can I Track Whether AI Recommends My Company?

Track a fixed set of real customer questions across relevant AI products, preserve the answers and citations, score mentions and accuracy with a disclosed rubric, and combine those observations with first-party platform reports, referral analytics, and qualified business outcomes.

Evidence-qualified answer · Updated October 17, 2026 · Measurement guidance checked against official OpenAI, Microsoft Bing, Google, and Anthropic documentation

Direct answer: You cannot reliably track AI recommendations with one vanity search or one universal rank. Build a reproducible prompt portfolio, standardize product and location conditions, record every valid response, inspect its citations, and measure mention rate, accurate mention rate, citation rate, query coverage, errors, referrals, and conversions separately. Always publish the tested scope and sample size.

Use four layers of evidence

Controlled answer audits

Run a stable set of customer questions under documented conditions. Capture which companies are named, how they are described, their order, qualifications, cited sources, and factual errors.

First-party platform reports

Use available reports from Bing Webmaster Tools and Google Search Console to observe citations, grounding contexts, cited pages, or generative-search impressions within those platforms’ supported coverage.

Website analytics

Track identifiable referrals and landing-page behavior. OpenAI says ChatGPT Search referral URLs include `utm_source=chatgpt.com`, but analytics will not capture answers that mention the company without sending a click.

Business outcomes

Ask leads how they found the company, preserve self-reported AI sources in the CRM, and track qualified inquiries, applications, purchases, or bookings without assuming every correlated visit was caused by a recommendation.

Step 1: Define what “recommended” means

Decide which answer events count before collecting results. A direct recommendation (“consider Company A for this need”) is different from a neutral mention, a warning, a citation to the company’s research, a navigational answer, or a linked product card. Mixing them creates an inflated and unusable score.

Event	Suggested classification	Record separately?	Reason
Company is explicitly suggested as an option for the stated need	Recommendation	Yes	Directly addresses selection intent
Company appears in a descriptive or comparison list	Mention	Yes	Visibility without necessarily advising selection
Company website supports a claim but company is not discussed	Citation	Yes	Measures source use, not brand recommendation
Answer provides an official link for a branded request	Navigational	Yes	Tests entity resolution rather than category preference
Company appears with a material warning or negative fit	Qualified or negative representation	Yes	High visibility may still signal risk or mismatch
Company is absent	No inclusion	Yes	Defines the denominator and prevents selective reporting

Step 2: Build the prompt portfolio

Start with questions customers genuinely ask, drawn from sales calls, support tickets, website search, market research, and public search data. Group them by intent so the measurement reflects the customer journey rather than hundreds of artificial keyword variations.

Branded: What does the company do? Where is it located? What are its current products or policies?
Category: Which providers solve the relevant problem for the intended audience?
Comparison: Which companies meet named criteria, and how do they differ?
Local: Which options serve a specific city, neighborhood, region, or service area?
Constraint: Which options meet budget, integration, accessibility, age, safety, policy, or compliance needs?
Research: Which companies or publishers provide credible evidence on the topic?
Post-purchase: Which source explains setup, support, cancellation, troubleshooting, or safe use?

Assign each prompt an importance weight only if the rationale is documented. Do not increase a prompt’s weight after seeing that it produces a favorable answer.

Step 3: Standardize the test

Product conditions

Record the product, visible model, plan, search or research mode, tool availability, and whether the answer used live web sources. Compare search-grounded modes with other search-grounded modes.

User conditions

Use a fresh conversation. Record signed-in state, personalization or memory settings, workspace controls, and connected sources. These factors may change retrieval or the generated answer.

Location conditions

State the relevant city, region, or country in local prompts. Record inferred or device location when known. Do not compare an Austin-localized result with a location-neutral result as if they are identical tests.

Save the exact prompt, not a paraphrase. Record the run date and time because models, indexes, listings, and public webpages change. When a request fails, hits a usage limit, or does not use the required search mode, mark it invalid rather than counting it as an omission.

Step 4: Preserve a complete evidence record

Field	What to save	Why it matters
Test identity	Prompt ID, intent group, importance, version	Keeps the denominator stable over time
Execution	Product, model, mode, plan, timestamp, success status	Separates unlike systems and invalid runs
Context	Location, account state, memory/personalization, fresh conversation	Documents factors that may alter the answer
Answer	Full response or durable capture, named companies, order, qualifiers	Allows later audit instead of relying on a summary score
Evidence	Every cited URL, source domain, page title, and source type	Shows which evidence supported that answer
Evaluation	Recommendation, mention, citation, navigational, absent, uncertain	Applies the predefined classification rubric
Quality	Materially accurate, minor issue, material error, unsupported claim	Prevents frequent misinformation from looking like success
Outcome	Referral visit, qualified action, self-reported source where available	Connects observable visibility with business value cautiously

Step 5: Calculate transparent metrics

Mention and recommendation rates

Mention rate = valid prompts naming the company ÷ valid prompts tested. Recommendation rate uses only answers that meet the predefined recommendation definition. Report both counts and percentages.

Accuracy rate

Accurate mention rate = valid prompts with a materially accurate company representation ÷ valid prompts tested. Also publish the number and severity of errors.

Citation rate

Owned citation rate = valid prompts citing an official company URL ÷ valid prompts. Track independent credible citations separately; they serve a different evidentiary role.

Intent coverage

Count distinct high-value intent groups with accurate inclusion. This prevents repeated variations of one easy branded question from overwhelming gaps in comparison, local, or category discovery.

Source coverage

Measure cited pages, credible domains, and source types. Diversity can reveal dependence on one fragile source, but more sources do not automatically mean better evidence.

Outcome rate

Track identifiable AI referrals and qualified actions. Keep these separate from mentions because an unlinked recommendation can influence a later direct visit that analytics cannot attribute confidently.

Always show sample size. A 100% recommendation rate based on two branded prompts does not mean the company dominates its category.

Use platform reports within their boundaries

ChatGPT referrals

OpenAI says ChatGPT Search automatically adds `utm_source=chatgpt.com` to referral URLs. Create an analytics segment for that source and review landing pages, engagement, conversions, and date trends. This measures identifiable visits—not every citation, mention, recommendation, or later direct visit.

Bing Webmaster Tools AI Performance

Microsoft’s current reporting can show citations, cited pages, grounding queries, intents, topics, trends, Citation Share, and comparison views across supported Microsoft and partner experiences. Microsoft states that Citation Share is observational, not a ranking system, traffic share, or quality score.

Google Search Console

Google’s generative AI performance report shows impressions for supported Google Search generative features, currently including AI Overviews and AI Mode, with page, country, and device views. Google’s documented click, impression, and position rules apply to links within these features. This is Google Search data, not cross-platform AI monitoring.

Claude citations

Anthropic says Claude web search retrieves multiple live sources and includes direct citations and source links. For a controlled audit, preserve those sources from each response. Anthropic does not describe this as a universal publisher dashboard, so answer-level auditing remains necessary.

Step 6: Establish a baseline before changing the site

Run the complete portfolio before a major intervention. Then document exactly what changed: a corrected listing, a new substantive service page, improved crawler access, updated pricing, consolidated duplicates, stronger evidence, or an earned third-party reference. Retest on the same cadence and conditions.

Do not change page structure, business profiles, prompt wording, test models, geography, and scoring rules simultaneously, then claim one tactic caused the result. AI systems and the web change independently, so most observational tracking shows association rather than proof of causation.

A useful change log contains the intervention date, affected URLs or profiles, reason, expected mechanism, validation result, and any indexing or crawl confirmation. Evaluate movement across several periods rather than announcing success from one regenerated answer.

Common tracking mistakes

Selective testing

Rerunning until the company appears
Publishing only favorable screenshots
Removing difficult prompts after the baseline
Counting failed runs as competitor wins
Changing the rubric after seeing results

False equivalence

Treating citations as recommendations
Calling answer order a documented rank
Combining model-only and web-search results
Comparing different locations without disclosure
Equating referral traffic with all AI influence

A sustainable reporting cadence

For a stable business, a monthly controlled panel is often more useful than daily manual checks. Active launches, rebrands, incidents, or rapidly changing product categories may justify weekly measurement. Preserve the core panel for trend continuity and maintain a smaller experimental panel for new questions.

Your report should show valid run count, missing tests, recommendation rate, accurate mention rate, citation rate, intent coverage, material errors, top cited pages, source domains, referrals, and qualified outcomes. Break out results by platform and intent. Include the methodology and version history with every report.

Browse the AI Tools directory and reviews of ChatGPT, Claude, Gemini, and Perplexity when selecting the products your audience plausibly uses.

PromptMarketMap can be linked as the related research and prompt-market layer, while RecommendationIndex can be linked as the related measurement product. Describe their results only within their published methods, coverage, and current feature set.

Minimum viable tracker

Prompts, conditions, answers, sources, accuracy

If you preserve those five things, you can audit and improve the program later. If you save only a score, you cannot explain what changed or whether the company was represented correctly.

Frequently asked questions

Can analytics show every AI recommendation?

No. Analytics can identify some referral visits, but an answer may mention the company without a link, and a user may return through search or direct navigation later.

Should I test from one account?

A consistent account can support trend tracking, but it may carry personalization or memory. Record those settings and consider a neutral test condition when customer personalization is not the subject.

Does being listed first mean I won?

Not necessarily. Generated-answer order is not a universal documented rank. Evaluate whether the company was actually recommended, the criteria used, factual accuracy, citations, and fit for the prompt.

Can I automate the tracking?

Yes, if the workflow preserves exact prompts, platform conditions, raw answers, citations, errors, and rubric versions. Automation should not hide missing runs or turn subjective labels into unexplained facts.

Official sources reviewed

Source review date: October 17, 2026. Platform reports, models, interfaces, and coverage change. Preserve raw evidence and verify current documentation whenever a trend is interpreted.

Track recommendations you can audit

A smaller reproducible panel is more valuable than a large opaque score built from unknown prompts and changing conditions.

Explore AI tools

Measure it

See what AI recommends in your market

RecommendationIndex is the measurement layer behind the portfolio. Use it to evaluate recommendation visibility across defined questions and AI systems instead of relying on one chat.

Check your company in RecommendationIndex →

Verified evidence · Reviewed October 2, 2026