How to Track Brand Mentions in AI Search

Unlike a search result page ranking, there is no fixed position to hold in AI search. If you run the same buying prompt twice in the same hour you can get different brands, different citations, different framing back. API-served models reproduce their own outputs in only 22.1% of tests, so it's most important to track frequency, not rank. That means a fixed panel of buyer prompts, run repeatedly across ChatGPT, Perplexity, Gemini, and Google AI Overviews, logged by hand first and automated only once you know what you're measuring.

Here's how to build that panel and turn it into a number leadership can actually track.

Why tracking AI search isn't SEO rank tracking

A rank tracker checks a stable artifact because Google crawls, indexes, and ranks, and the position it reports today is roughly the position every searcher in that market sees.

An AI engine generates its answer at request time, per session, with no persistent index of results to poll. Re-run the same prompts weeks later and the citations behind the answers routinely shift, on some engines more than others.

Instead of asking what position you hold for a keyword, you're asking what share of answers you appear in, how you're described, and which sources put you there. Your existing rank tools can't answer any of that. They track keyword positions on SERPs, and there's no SERP to track here. Queries are conversational prompts with no fixed universe, and answers vary run to run, so measurement means sampling the same prompts repeatedly over time.

The four platforms to monitor first

Each of the four major surfaces selects sources differently, so a mention on one tells you almost nothing about the others. Only a small fraction of the top sources cited across ChatGPT, Perplexity, and Google AI Overviews actually overlap.

• ChatGPT: Leans on authoritative reference sources, with Wikipedia dominating the top results even though it's a modest share of total citation volume. Watch it for category-definition and comparison prompts.
• Perplexity: Cites visibly, with numbered sources, which makes citation logging easiest here. YouTube and Reddit lead its citation mix by a wide margin.
• Gemini: Mentions brands generously in prose but attaches citations sparingly, so log mentions and citations as separate events. It passes a referrer (gemini.google.com) when users click through, which matters for analytics later.
• Google AI Overviews: Leans on user-generated content far more than the other three engines, pulling heavily from forums and reviews. It sits inside the largest search surface, so absence here costs the most impressions.

Track all four separately from day one. A blended view can hide the platform where performance is weakest.

Building your prompt panel

Once you know where to look, the next job is deciding exactly what to ask. Write prompts in buyer language, as full questions with use cases and constraints. Then freeze the set. Week-over-week numbers only mean something if the inputs don't move, and one widely used measurement framework recommends a frozen prompt set of 30 to 60 prompts per topic cluster, re-run weekly. Start with 20 to 30 per product line across five prompt types:

• Category discovery: "What's the best [category] software for a [company profile]?" This measures unaided presence at the top of the funnel.
• Comparison: "Compare [your brand] and [competitor] for [use case]. Which should I pick?"
• Alternatives: "What are the best alternatives to [competitor]?" Run it against your own name too. The response shows which competitors the engine associates with your brand.
• Validation: "Is [your brand] worth it for [ICP]? What are the drawbacks?"
• Problem-first: "How do I [problem your product solves]?" No brand named. This tests whether engines surface you from the pain alone.

Pull the phrasing from sales call transcripts and support tickets where you can. A panel built from marketer vocabulary measures a conversation buyers aren't having.

The manual tracking workflow

We push every team we work with to run this by hand first, before any tool. The manual pass teaches you what the dashboards will later summarize, and a spreadsheet is all you need to start.

1. Run prompts as a real buyer. Use a clean browser session, logged out where the platform allows it, with no conversation history. Personalization skews answers toward what the account has already seen.
2. Repeat each prompt three to five times per platform. One run gives you noise. Repeated runs give you a measurement, because outputs vary so much between identical requests.
3. Paste the full raw answer into your log. Not a summary. Framing analysis later depends on exact wording.
4. Record structured fields alongside it. Record date, platform, prompt ID, and run number. Then record brand mentioned (Y/N), position in any list, competitors named, sentiment, the exact framing phrase, and every cited domain.
5. Roll up weekly into mention rate and share of voice per platform.

Metrics to track and how to report them

Every logged run tells you whether you appeared, how much of the conversation you own, how you were framed, and who the answer cited.

Mention frequency and share of voice

Mention frequency is the share of prompt runs where your brand appears at all. AI share of voice compares you against competitors. A simple formula is "(number of my brand mentions / total number of all brand mentions) × 100," counting each brand once per response even if it's named twice.

Worked example. 25 prompts run 4 times each is 100 responses. Your brand appears in 18, Competitor A in 30, Competitor B in 12. Total mentions come to 60, and your AI share of voice is 30%.

Report it per platform. A blended number can hide that you own Perplexity and are invisible in AI Overviews, and frankly, that's the kind of gap that costs you the board's confidence once someone else notices it first. Since no industry-standard formula exists and vendors weight position and topic volume differently, state your formula in the report once and never change it mid-trend.

Sentiment and framing

Log a positive, neutral, or negative call plus the exact clause around your brand name, because "a budget option for small teams" and "the standard for enterprise" are both mentions, and they're different assets. One study of 102 brands across 102,025 responses found presence flipped only 6.8% between measurements while sentiment framing flipped 45.5% of the time. Counting mentions without reading framing understates your risk by an order of magnitude.

Citations and source analysis

Log every domain each answer cites, and log citations separately from mentions, because the two diverge sharply. One citation study found ChatGPT rarely names a brand outright but links to its domain in most of the responses where it does appear. Gemini runs the opposite way. It mentions brands generously in prose but attaches a source link only a small fraction of the time.

One analysis of 6.8 million citations found brand-controlled sources drive 86% of citations, with first-party websites accounting for 44% and listings for 42%. Your site and listings are the base you control directly. The third-party remainder, the review sites and publications the engines keep citing in your category, becomes the lever list in your citation log, and your PR or partnerships team uses it to close mention gaps from outside.

Running competitor analysis in AI answers

You've been logging your own numbers on this panel. Turn it on your competitors next by running the identical panel with competitor names substituted into the branded prompts. The comparison and alternatives prompts already capture most of it, so this is largely a second read of data you've logged. Compute each competitor's share of voice with the same formula.

If your share of voice and Competitor A's both drop 10 points in a month, the model's behavior changed. If only yours drops, you have a problem. And log which domains get competitors cited. Competitor-only citation domains become named, addressable outreach targets instead of vague "do more PR" line items.

Tying AI visibility to traffic and revenue

Visibility numbers matter most once they connect to what the business already tracks. Track crawler activity in your server logs and human referrals in GA4. In server logs, hits from OAI-SearchBot, which surfaces websites in ChatGPT search results, and PerplexityBot confirm the engines are fetching your pages at all.

In GA4, know the blind spots first. The built-in AI Assistants channel excludes AI Overviews and AI Mode, which GA4 keeps under Organic Search because they run on Googlebot with no separate referrer. GA4 cannot distinguish the largest AI search surface from ordinary organic clicks. ChatGPT's mobile app passes no referrer either, so GA4 classifies that traffic as Direct.

Blind spots to label in the report:

• GA4 keeps AI Overviews and AI Mode under Organic Search.
• GA4 often classifies ChatGPT mobile as Direct.
• Visible AI referrals undercount influence.

For what GA4 can see, build a custom channel group. Go to Admin → Data display → Channel groups, add a channel with the condition Source matches regex, and move it above Referral so the sequential waterfall catches it. A ready-made regex covers the major sources:

^https:\/\/(www\.meta\.ai|www\.perplexity\.ai|chat\.openai\.com|claude\.ai|chat\.mistral\.ai|gemini\.google\.com|bard\.google\.com|chatgpt\.com|copilot\.microsoft\.com)(\/.*)?$

Expect small volume and outsized quality. One ecommerce dataset shows ChatGPT referrals converting at 11.4% versus 5.3% for organic search. GA4 referral counts undercount the effect too. Users who get an AI recommendation are 2.5 times more likely to visit the brand's site within seven days, and over half of those visits arrive through branded search, not a direct AI referral. Put branded search volume on the same report as AI share of voice to catch the lift GA4 misses.

Tools that automate AI mention tracking

Manual logging teaches the method. At some point, though, the panel outgrows a spreadsheet, and a tool earns its fee by doing what you can't do by hand. It runs the panel daily at sample sizes that smooth out non-determinism, computes share of voice, extracts every citation, and keeps trend history. It does not pick your prompts, judge whether "cheap option" is acceptable framing, or produce the content that closes a gap. That work stays with your team.

• CheckThat (CheckThat.ai, freemium): The free tier lets you browse brand visibility across 5,800+ brands and 1,900+ B2B software categories built from 2.6M+ AI responses, no credit card required. The premium workspace tracks up to 50 custom prompts plus 100,000+ industry prompts, with sentiment monitoring and daily historical responses across ChatGPT, Claude, Gemini, and Perplexity. We built it.
• Otterly.ai (Lite $29/mo, 15 prompts): It runs daily tracking on every plan and queries as a neutral, non-personalized user so browsing history doesn't skew results.
• Profound (Starter $99/mo, 50 prompts on ChatGPT): It captures responses directly from the consumer browser experience rather than API outputs. Enterprise covers up to 10 engines.
• Semrush AI Visibility Toolkit ($99/mo per domain): It draws on a prompt database of 289M+ prompts updated daily and shows AI share of voice next to SEO share of voice in one dashboard.

Keep your manual spreadsheet running for the first month after adopting any tool. Comparing its numbers to your own logs shows you how its sampling differs from a real buyer session, and where to trust it.

Building a repeatable monitoring cadence

Whether you're running this by hand or through a tool, someone still has to own the rhythm. Assign one owner. In most teams that's the content or demand gen lead who built the panel. Split the rhythm into three loops.

Weekly: The owner runs the panel (or reviews the tool dashboard), updates per-platform share of voice, flags sentiment flips with the quoted framing, and notes new or lost citation domains. Budget 30 to 60 minutes with tooling, or a few hours without it.

Monthly, for leadership: One page. Share of voice trend per platform against your top two competitors, a sentiment summary with two or three quoted framings, citation domain movement, AI referral sessions and conversions from the GA4 channel group, and actions taken plus actions planned. Never report a single week's swing as a win or a crisis. Use four-week trends at minimum, given how much answers churn.

Quarterly: Refresh the panel for new competitors, products, and buying language. Version the change and annotate the trendline so nobody compares pre-refresh and post-refresh numbers blindly.

We've watched this exact bottleneck stall more than one otherwise-good tracking program. If the weekly run starts crowding out the content work it's supposed to inform, have the owner consolidate the tracking, content, and SEO workflows into a single system rather than running them separately.

Closing your AI brand-mention gaps

Tracking tells you where you stand. Closing the gap is the work we spend most of our time on with clients, and it starts with earned third-party mentions. A 75,000-brand study found brand web mentions correlate with AI Overview visibility at 0.664, well ahead of Domain Rating (0.326) and total backlinks (0.218), and brands in the top quartile of web mentions receive up to 10× more AI Overview mentions than the next quartile. A follow-up piece in Search Engine Journal covered the same brand-mention correlation. Point your digital PR at the citation log you built. The review sites and industry publications each engine already cites in your category are named there, and YouTube channels count too when they show up in the log.

Citation work queue

Use the citation log as a work queue:

• First-party fixes: pages, listings, directories, and review profiles you control.
• Third-party targets: review sites and publications that cite competitors while skipping you.

First-party content fixes

On your own site, the moves with the strongest evidence:

• Answer-first structure: Lead each section with the direct answer, use clear heading hierarchy, and format comparisons as tables and Q&A blocks. Engines lift self-contained passages, so pages built as extractable chunks travel further than essays.
• Quotable evidence: The original GEO research (KDD 2024) found adding statistics, quotations, and source citations were the top-performing optimization methods, boosting visibility up to 40%, while keyword stuffing performed 10% worse than baseline on Perplexity. Publish original numbers other people will cite.
• Entity consistency: Describe your brand identically across your site, listings, directories, and review profiles. LLMs learn brand associations from co-occurrence in text, and five slightly different one-liners dilute the entity you're trying to establish.

Hallucinated mentions, where an engine states something false about your product, will also show up in your logs. The major platforms don't document a dedicated factual-correction workflow for brands. Perplexity accepts reports through the flag icon or support@perplexity.ai, routed through Perplexity support with the query URL, a description of the error, and the expected result. Google offers a Report a problem link beneath each AI Overview. OpenAI's trademark dispute form covers infringement only, not factual errors. File the reports, then treat the durable fix the same as the visibility fix. Publish accurate, structured pages the engines can crawl, and get the correct version echoed on the third-party domains they cite.

That same discipline, publish once and let the citations follow, is what turns a monitoring habit into an actual visibility program. GrowthOS runs the loop end to end. Insights tracks your citations and share of voice across the same engines you're already logging by hand, and feeds each gap straight into Creation so the content that closes it gets made next. If your tracking spreadsheet needs a system behind it, book a demo. Engagements start from $6,000/mo.