# How Enterprises Measure and Improve Brand Visibility in AI Answers > A practical routine for tracking mentions, citations and accuracy in ChatGPT, Perplexity and Google AI Overviews, with AEO and GEO next steps. Source: https://terradium.io/blog/measure-brand-visibility-in-ai-answers **Rashi Lachuriya (guest)** Last updated 5 November 2026 ![Hero image: the title "How Enterprises Measure and Improve Brand Visibility in AI Answers" beside logos of AI answer platforms](https://terradium.io/media/guides/measure-brand-visibility-in-ai-answers.webp) *How to track mentions, citations, accuracy and trust across ChatGPT, Perplexity and Google AI Overviews* A buyer opens ChatGPT and asks which vendors are worth a look. The shortlist may be settled before anyone visits a results page. If your company isn't on it, or is on it with the wrong description, nothing in your usual reporting will say so. This isn't the same problem as search rankings. A rank tracker tells you where a page sits for a keyword, and that position holds still long enough to chart. An AI answer can change every time the question is asked, so one screenshot proves very little. What you need is a pattern: the same questions, asked again and again, across several platforms and separate sessions. ## Key takeaways - One AI response is an anecdote. Track how often your brand appears across a fixed set of prompts, each run several times. - Measure mention rate, citation rate, source mix, accuracy and coverage across the buying journey. Skip any "average position" inside an answer, because there is no stable position to average. - Outside sources carry weight. Reviews, trade coverage, directories and original research all shape what AI systems pull in. - The published before-and-after cases come from a vendor and start from near-zero baselines. Read them as a method worth testing, not a result to expect. - A wrong price, feature or security claim in an AI answer can cost you a deal before anyone talks to you. Give monitoring a named owner. - The first cycle can be small: a prompt list, a spreadsheet and one person responsible. ## Building a tracking routine across ChatGPT, Perplexity and Google AI Overviews ### Start with a fixed list of buyer questions Aim for roughly 30 to 40 prompts and lock the wording. If you rephrase between rounds, you can't tell whether the AI changed or you did. Spread them across the way people actually buy: - **Discovery:** "What are the best \[category\] tools for \[use case\]?" - **Comparison:** "How does \[brand\] compare with \[competitor\]?" - **Evaluation:** "Is \[brand\] a good fit for a company that needs \[specific requirement\]?" - **Branded research:** "What does \[brand\] offer?" or "What should buyers know about \[brand\]?" Write them in buyers' words, not yours. Sales-call notes, support tickets, customer interviews and Search Console queries make better raw material than your own messaging, which describes the company the way the company likes to be described. The question to keep asking is which of these could plausibly shape a real shortlist. ### Run every prompt more than once AI answers don't sit still the way a results page does. Ask the same question twice and the brands, the sources and the wording can all shift. SparkToro co-founder Rand Fishkin and Patrick O'Donnell of Gumshoe.ai tested this with 600 volunteers who ran 12 prompts across ChatGPT, Claude and Google's AI features, producing 2,961 responses ([SparkToro's write-up](https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/)). Each prompt ran 60 to 100 times per platform, and the chance of two runs returning the same brand list was under 1 in 100. [Search Engine Journal](https://www.searchenginejournal.com/ai-recommendations-change-with-nearly-every-query-sparktoro/566242/) summarized the results and noted that the study isn't peer reviewed and that Gumshoe.ai sells AI-tracking tools. Keep both facts in mind. The part that matters for your routine is a secondary finding. How often a brand showed up across many runs was steadier than where it landed in any single answer. That is why this process counts appearances and ignores ranks. As a starting point, run each prompt five times on each platform, in fresh sessions with the same settings every time. Five is our practical suggestion, not an established threshold. The SparkToro runs were far larger and nobody has settled how many are enough, so go higher for the prompts that matter most. ### Log mentions and citations separately They sound alike but they're different events. A mention means the answer names your brand. A citation means it links to one of your pages or lists your site as a source. You can be named without being cited, and an answer can lean on third-party pages instead of your own. For every run, record the date, platform, mode, prompt, run number, whether the brand was mentioned, whether your domain was cited, and which other sources appeared. Give Google AI Overviews and AI Mode their own columns. [Google says](https://developers.google.com/search/docs/appearance/ai-features) these features can use different models and techniques, so the answers and links may not match. ### Decide whether doing this by hand is enough Manual checking is free, and you see exactly what each platform returned. The catch is volume. Thirty prompts on three platforms at five runs each comes to 450 responses to read and log. If you evaluate a tool, ask how many runs it performs per prompt, whether the method is documented, and how it handles personalization and session differences. Be wary of any dashboard that reports a precise "rank" inside AI answers without showing the math. Answers vary too much from run to run for one position to describe them. Google adds a related warning in its [guide to generative AI search](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide): no third-party tool has access to its internal ranking or AI systems. ## Which metrics are worth tracking Think in rates and quality checks, not a leaderboard. **Mention rate** is the share of relevant runs in which your brand appears. Split it by prompt type (discovery, comparison, evaluation). Strong results on branded questions can hide a weak showing on discovery, which is where new buyers come from. **Citation rate and source mix** show how often your own domain is cited and which third-party sites appear alongside it in answers about your category. The source mix tells you where to put content and outreach effort. **Referral traffic and branded search** can back the picture up. Both tend to move after your visibility changes, not before. **Search Console** now gives you a first-party view. Google [announced its generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports) on June 3, 2026, and a note on that post says they reached all websites worldwide as of August 31, 2026. The Search report covers AI Overviews and AI Mode, with impressions by page, country and date, plus a device breakdown. It counts impressions, so it shows how your pages appeared and says nothing about how an AI answer describes your brand. Clicks from AI features stay in your regular Performance report, [mixed in with the rest of your search traffic](https://developers.google.com/search/docs/appearance/ai-features). Some properties may still see no report, particularly sites with little AI activity, [according to Search Engine Journal](https://www.searchenginejournal.com/google-search-console-ai-reports-rolled-out-worldwide/587836/). Either way, keep your own log split by feature. ### Metrics that mislead - An "average position" inside an AI answer, as if it were a stable ranking - One screenshot of a flattering answer - Mention counts with no note of how many prompts and runs were tested - An "AI visibility score" whose method isn't disclosed - Citation counts that blend branded and discovery prompts, since branded questions naturally lead to your own site ## What can improve your visibility in AI answers Nobody can guarantee that an AI system will name or cite your company. What you can do is improve the quality and availability of the information those systems might draw on. - **Describe yourself the same way everywhere.** State your category, products, locations and ideal customers plainly, and keep that wording consistent across your site, company profiles, directories and press material. If one source calls you a software provider and another calls you a consultancy, the picture an AI assembles will be muddier. - **Show up where the answers already look.** Your source-mix log will show which review platforms, trade publications, comparison pages, directories and forums keep appearing. That becomes a short outreach list, and it beats chasing mentions on sites your buyers never visit. Aim for real coverage. Google's guide says seeking [inauthentic mentions](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) isn't as helpful as it might seem. - **Publish claims others can check and quote.** A [research paper on generative engine optimization](https://arxiv.org/abs/2311.09735), presented at [KDD 2024](https://doi.org/10.1145/3637528.3671900), tested several content methods on a benchmark of queries. Its authors reported visibility gains of up to 40% from approaches such as adding citations, quotations and statistics, with results that varied by domain. These were controlled experiments on generative engines as they stood then, so they don't promise the same lift for every business. The practical lesson is plainer: make specific claims, back them with credible evidence, and say where your numbers come from. - **Put the answer first.** Nobody should wade through a long introduction to find it. Lead with the direct response, then give the reasoning, the limits and the next steps. Your prompt set doubles as a content backlog here. Questions that keep returning incomplete or wrong answers may point to gaps on your own site. - **Make sure crawlers can reach your pages.** Good content does nothing if the systems that retrieve it can't get in. Google says [no special files or markup](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) are needed to appear in its AI features. For ChatGPT search, [OpenAI's crawler documentation](https://developers.openai.com/api/docs/bots) separates OAI-SearchBot, which surfaces sites in search results, from GPTBot, which relates to model training. Blocking one doesn't mean you have to block the other. Compare your robots.txt and crawler settings with what your business wants before you change anything. ## Two published before-and-after cases from one vendor General advice is easy to nod along to, so here are two named examples with the numbers Profound reported. Both come from case studies published by Profound, an AI-visibility vendor. That makes them self-reported and not independently audited. ### Case 1: Ramp, accounts payable software **Prompt set.** Profound doesn't publish Ramp's prompt list or run counts. Its [case study](https://www.tryprofound.com/customers/ramp) frames the gap as low visibility for accounts payable, among fintech brands. The titles of the pages Ramp built hint at the shape of the questions: accounts payable software for small businesses, for large businesses, and automation tools. That is our inference, not something the page states. **What changed.** Profound's analysis found AI engines often cited content about automation, AI and software comparisons, and says about 6% of Ramp's existing citations came from that kind of content. Ramp then published two pages named for the exact buyer question, *Accounts Payable Software for Small Businesses* and *Accounts Payable Software for Large Businesses*, along with comparison pages such as *Top 6 Accounts Payable Automation Software*. **Before and after.** The headline figure is AI visibility rising from 3.2% to 22.2% within a month, with Ramp moving from 19th to 8th among fintech brands in that category. The two new pages drew more than 300 citations in a month. The chart on the same page tells a more modest story. It compares Dec. 1, 2024 with Feb. 15, 2025, a window of about 11 weeks, and shows 22.2% visibility with a 12.2-point change, which implies a starting point near 10%. Profound's page also lists a citation share of 8.1% for Feb. 15, 2025, but it doesn't label the starting figure clearly, so treat it as a separate measure from visibility and not as a before-and-after result. ### Case 2: 1840 & Co., remote staffing **Prompt set.** Profound's [case study](https://www.tryprofound.com/customers/1840-co-answer-engine-optimization-case-study) says 1840 & Co. scanned hundreds of queries across AI engines, including "best remote staffing agencies" and a similar search for U.S.-based developer staffing. The company started at 0% visibility, against roughly 74% for Toptal and 29% for Upwork, as reported by Profound. **What changed.** The company published one blog post, *10 Best Agencies For Remote Staffing Solutions*. It copied the formats Profound saw being cited: clear headings, short comparison lists, the company placed near the top, and brief FAQs. **Before and after.** Visibility went from 0% to 6% in two weeks and to 11% by the end of the month, which Profound says made it one of the top five brands cited for that space. ### What the cases do and don't show Both brands started from low visibility on unbranded questions, both published pages written in the buyer's own wording, and both measured the same prompts again afterward. That is the part you can borrow. The limits matter just as much. The vendor defines the visibility metric, the prompt lists and run counts aren't public, and climbing from near zero is easier than improving a brand that already appears in most answers. Treat both as hypotheses to test with your own prompt set, and judge any change over weeks rather than from a single reading. ![Infographic: Brand Visibility in AI Answers, with four columns (track it, measure it, improve it, keep it accurate) summarizing the routine described in this post](https://terradium.io/media/guides/measure-brand-visibility-in-ai-answers-infographic.webp) ## When an AI answer gets your brand wrong Wrong answers aren't hypothetical. In March 2025, Minnesota solar installer Wolf River Electric sued Google in state court, alleging that an AI Overview falsely said the state attorney general was suing the company and that customers canceled contracts after seeing it ([Futurism reported the allegations](https://futurism.com/company-sues-google-ai-overviews)). Google had the case moved to federal court. In January 2026 a federal judge [sent it back to state court](https://storage.courtlistener.com/recap/gov.uscourts.mnd.225722/gov.uscourts.mnd.225722.34.0.pdf), ruling that the removal came too late. The underlying claims haven't been decided, so the case doesn't prove the allegations. It does show what a wrong answer can cost. For most enterprise and SaaS companies the damage is more everyday: a wrong price or plan limit, an integration you don't offer, an outdated security claim, a skewed comparison with a competitor. A buyer can act on that before ever contacting you. Keep a short list of facts that must be right, confirmed by product and by legal where it applies. Log every wrong answer with the date, platform and prompt, trace it to the page it came from, fix that page, and use the platform's feedback channel if there is one. Then name one owner. Without one, marketing sees the answer, legal owns the response, and nobody is watching. ## A 30-day measurement plan ### Week 1: set the baseline Build the prompt set and the list of facts that must be correct. Assign an owner and set up a simple tracking sheet. ### Week 2: run the checks Test each prompt five times per platform, in fresh sessions with consistent settings. Record mentions, citations, sources and factual errors. ### Week 3: find the biggest gaps Compare mention rates by funnel stage and note which sources are cited most. Choose three actions, such as improving a key page, correcting a directory listing or approaching a relevant publication. ### Week 4: make changes and retest Publish the updates, rerun the same prompt set and compare overall rates and accuracy. Don't fixate on individual answers. Expect noise. A small sample moves on its own, so a minor swing from one month to the next means little. Look for larger, repeatable shifts, and keep the method identical each round. After a few months that gives you a far better trend line than a folder of unrelated screenshots. ## Frequently asked questions ### Can we measure AI Overviews in Google Search Console? Partly. Google [announced the generative AI performance reports](https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports) on June 3, 2026, and says they reached all websites worldwide as of August 31, 2026. The Search report shows impressions in AI Overviews and AI Mode by page, country and date, with a device breakdown. It counts impressions only. Clicks from these features sit in your regular Performance report, and nothing here tells you what an AI answer says about your brand. Some sites, especially those with little AI activity, may not see a report. ### Does generative engine optimization (GEO) replace SEO? No. Google's own [guidance](https://developers.google.com/search/docs/fundamentals/ai-optimization-guide) treats optimizing for its generative AI features as SEO, and the advice it gives is the familiar kind: crawlable pages, helpful content, clear structure. Thinking about AI answers is one more reason to make content accurate, specific and well supported. It isn't a replacement for the fundamentals. ### What should we do if an AI answer describes our company incorrectly? Save the answer with the prompt, platform and date. Find where the claim likely came from and correct that source if you can, then use the platform's feedback channel. If the error touches pricing, security or contract terms, bring in legal. Don't expect the correction to show up right away. ## Final thought AI visibility will probably never be as tidy as a rank report. Answers shift, sources change, and a mention alone doesn't tell you whether what was said is right. A steady routine can still show whether you appear in the right conversations, which sources shape them, and where your team should focus next. Start with a prompt set you can manage, write down what you find, and let each round decide the next one. ## About the author **Rashi Lachuriya** Rashi Lachuriya works in marketing at [Samta.ai](https://samta.ai/), an enterprise AI consulting firm that helps companies move AI projects from pilot to production. She writes about AI strategy, governance and enterprise AI adoption.