How to track AI brand recommendations and qualify cited sources
A repeatable workflow for measuring brand mentions, recommendations, and citations across ChatGPT, Gemini, and Perplexity, then choosing a useful next action.
How do you track how ChatGPT, Gemini, and Perplexity recommend your brand?
Ask a fixed set of buyer questions repeatedly in each assistant, save the answers and supplied citations, and classify mentions separately from recommendations. Report each engine with its own successful-capture denominator. Inspect the cited pages before choosing a content update or editorial inquiry. Siftree AI visibility workflow is a starting workflow for this work; the collected answers show what was said and cited, not every source an engine used or why it selected a brand.
How this workflow fits: Siftree is an AI-native IT department for your team. Ask for a ai visibility solution through your existing AI connected over MCP. Siftree builds and runs the solution within your budget, managing integrations, models, databases, hosting, and access controls. Explore the sources and connections for your workflow through the available integrations.
Start with the buying decision you want to measure
An SEO or brand lead needs to know whether an assistant recommends the product for a relevant job, how it describes the product, and what evidence the answer gives. Those are separate questions. A brand can be named in a warning, cited without a recommendation, or recommended without a link.
Siftree AI visibility workflow brings scheduled buyer questions, captured answers, brand labels, cited sources, and product-fact checks into one workflow. Confirm collection access, engine coverage, and any CMS write permissions when scoping the app. A custom publishing action needs the appropriate connected system and review rules.
Use the same questions for the baseline and follow-up. Keep unbranded questions such as “What tools track AI brand recommendations?” separate from branded questions such as “Does our product support Gemini?” A branded prompt is useful for checking accuracy but does not measure discovery in the same way.
Preserve enough context to check an answer
Keep the exact prompt, engine and available model identifier, capture time, locale, collection method, search mode, raw answer, supplied citation URLs, and capture status. Record whether the answer came from a consumer interface or an API; treat different collection methods as different cohorts. Do not substitute one for another in a trend without labeling the change.
A failed request is a collection failure. A successful answer with no links is an answer without citations. Keep those states distinct. Save the original answer before extracting labels so a reviewer can check whether a supposed recommendation is actually a comparison or a negative mention.
Define the metrics before reading the trend
| Metric | Counting rule |
|---|---|
| Mention rate | Usable answers naming the brand, divided by usable answers in the same engine and prompt cohort |
| Recommendation rate | Usable answers affirmatively suggesting the brand for the requested job, divided by usable answers in that cohort |
| Owned citation rate | Usable answers linking to at least one page on your site, divided by usable answers in that cohort |
| Source-page presence | Distinct usable answers citing a particular normalized page, divided by usable answers in that cohort |
| Capture coverage | Successful captures divided by planned captures; show failed and missing captures separately |
Count a source page once per answer for source-page presence. Remove text-fragment anchors and consolidate verified URL variants, but keep genuinely different pages separate. Preserve original citation URLs alongside normalized ones. Several citations to one article in a single answer do not constitute several independent answers.
Work through a small example
The answer-visibility calculations trace how mentions, recommendations, and citations are counted across four answer records.
Three captures succeed and one fails. Two of the usable answers name the brand; one recommends it. One links to the brand's own page. The resulting mention rate is 2/3, recommendation rate is 1/3, owned citation rate is 1/3, and capture coverage is 3/4. Two fragments of the same publisher article in one answer count as one answer citing that page. Another usable answer contains no citations and still belongs in the mention-rate denominator.
Manually check the recommendation label against its passage before reporting a win. Keep the fractions visible rather than presenting a precise-looking percentage from three answers.
Turn the source map into a decision
First inspect the actual cited page. Is your brand already included? Is the article still relevant? Who publishes it? Is there a legitimate review, submission, or correction process?
| Evidence you find | Useful next step |
|---|---|
| Your page omits a supported capability | Add a clear explanation, requirements, limitations, and a worked example to that page |
| A publisher evaluates several relevant tools | Offer verifiable product facts and ask about its evaluation process |
| An integration directory has a qualifying category | Confirm eligibility and document the integration before submitting |
| A competitor cites its own product page | Study how it answers the buying question and improve your own explanation |
| The answer contains no citations | Review the answer and your product coverage; mark its underlying sources as unknown |
A supplied citation is evidence that the answer linked to a page. It does not prove the page caused a particular claim. A citation map also does not establish whether that page links back to you. Check those links independently before describing an opportunity as a missing backlink.
Repeat the measurement after the change
Log the page changed, publication date, and any external coverage. Rerun the fixed questions with matching collection settings and comparable repetitions, reporting each engine separately. Retain questions with no improvement and failed captures rather than silently dropping them. A small set can reveal examples to investigate; it cannot establish a universal ranking or isolate the effect of one link.
For Google AI Overviews and AI Mode specifically, Google's guidance emphasizes normal search eligibility, accessible text, and helpful content. It does not require a special AI text file or schema. Other chat products need their own discovery checks.
Review the answers behind any material change, then compare qualified visits and demo requests where attribution is available. A stronger recommendation trend is useful when it reflects accurate product fit and helps buyers take the next step.
Sources & methodology
Product details and supporting references used in this guide.

Co-founder, Siftree
Kyle co-founded Siftree after working in data across telecom, media, insurance, automotive, and entertainment and tourism. He writes Siftree's software comparisons and buying guides from the vendor documentation linked in each article.
See it with your own workflow.
Start with AI visibility, or have us build around your team's workflow. Bring the data you use, the work you want done, and where the result should go. We'll scope the app and required access with you.