Semrush's 2026 AI Visibility Index found citation overlap below 50% for every pair of AI platforms it tracked, which means being cited in one engine tells you little about the next. (Semrush report, PDF) That alone should cure you of trusting a single screenshot. AI search visibility is how often, and how accurately, AI assistants and AI search features name or cite your brand when buyers ask the questions that lead to a purchase.
Measuring it is awkward. There's no rank to check, answers change from one run to the next, and most of the influence never shows up as a click. Plenty of teams either skip it or screenshot one flattering answer and call it a trend. Both are mistakes. Our view: a hand-run baseline, repeated monthly against a fixed prompt set, beats any tool you buy before you've read the answers yourself.
This guide sets out a tracking setup a B2B SaaS team can run every month: which prompts to track, which four numbers to record, how to handle the noise, which first-party data to pull, and how to connect it to pipeline without overclaiming. If you're new to the topic, start with the pillar on what AEO is.
The four numbers worth recording
The engines to track are usually ChatGPT, Perplexity, Gemini and Google's AI Overviews and AI Mode, plus any others your buyers use. A tracking sheet with twenty columns gets abandoned. For each prompt and each engine, these four are enough:
Metric | Question it answers | How to record it |
|---|---|---|
Mention | Did the answer name us? | Yes / no |
Citation | Did it link to or cite one of our URLs? | Yes / no, plus which URL |
Accuracy | Did it describe us correctly? | Correct / partly wrong / wrong, with a note |
Position | Where in the answer did we appear relative to competitors? | First named, listed, or absent |
Mention and citation need separate columns because engines treat them differently. In Semrush's "ghost citations" study, 61.7% of citations came with no brand mention in the answer text, and Gemini named brands in 83.7% of appearances but linked them only 21.4% of the time. (Semrush)
Accuracy matters to sales. Being named with the wrong pricing model, a missing integration or a competitor's feature can do more harm than being left out.
From these rows you can work out share of voice: the share of tracked prompts where you're mentioned, compared with each competitor on the same prompts. SEO consultant Matt Diggity listed mentions, citations, estimated impressions and AI share of voice against competitors as the metrics to benchmark in a November 2025 post on X. (Matt Diggity on X)
Step 1: Build a buyer prompt set
The prompt set decides whether the whole exercise means anything. Track the questions a buyer asks near a decision. A prompt like "What is [your brand]?" will almost always mention you and tells you little.
Start with 30 to 60 prompts, grouped by intent:
Category shortlists: "best [category] software for [segment]", "top [category] tools for a [size] team".
Alternatives: "alternatives to [competitor]", "[competitor] vs [you]".
Fit and integration: "does [you] integrate with [tool]", "[category] tool that works with [stack]".
Problem-led: "how do I [job the product does] without [pain]".
Pricing and trust: "how much does [you] cost", "is [you] SOC 2 compliant" (only if that's true and documented).
Source the wording from real buyers. Sales call notes, demo form free-text answers and support tickets are better than a keyword tool here, because people phrase questions to assistants in longer, more conversational ways. Ahrefs suggests asking customers on lead forms or surveys which AI prompts they used to find you. (Ahrefs)
Freeze the set once it's built. If you change prompts every month, you can't compare months. Add new prompts in a separate group and keep the original set as your baseline.
Step 2: Pick engines and a sampling cadence
Track the engines your buyers use. For most B2B SaaS companies that means ChatGPT, Google AI Overviews and AI Mode, Perplexity and Gemini. Ask on sales calls if you're unsure.
Don't assume one engine stands in for the others. Across 126 million U.S. prompts (January to April 2026), Semrush's 2026 AI Visibility Index counted roughly 15 sources per ChatGPT response against about 3 for Gemini. (Semrush)
A cadence that works for most teams:
Run the full set at least once a month, in a logged-out or fresh session, from the country you sell in.
Run your 10 to 15 most important prompts weekly. A monthly snapshot alone is too coarse to separate a real change from run-to-run noise.
Run each prompt more than once per engine (three runs is a reasonable start) and record how many runs mentioned you.
Note the date, the engine, the model or mode where it's visible, and the location.
Step 3: Treat single answers as noise
The same engine gives different answers to the same prompt. In a Semrush study with Kevin Indig, GPT-5.2 ran the same 100 prompts in minimal and high reasoning modes, and only 25.6% of cited domains overlapped between the two. (Semrush)
So a single screenshot proves very little, in either direction. Three habits keep the data honest:
Report rates, such as "mentioned in 2 of 3 runs on 18 of 40 prompts", instead of yes or no per prompt.
Look at three-month trends before you change strategy.
Mark known model changes on the chart. After a ChatGPT update on May 7, 2026, Similarweb saw about 60% of ChatGPT referrals land on homepages, up from roughly 26 to 32% before. (Similarweb) A shift like that can move your numbers without anything changing on your site.
Step 4: Pull the first-party data you do have
Prompt sampling is an estimate. Two first-party sources sit alongside it.
Google Search Console
Google's Generative AI performance report, launched in June 2026, shows impressions from AI Overviews and AI Mode with page, country, device and date breakdowns. (Google Search Central Blog) It's the only data that comes from inside one of the engines. Use it to see which of your pages appear in Google's AI features and whether that's growing.
Check one setting while you're there. Per Google's optimization guide, the site has to be opted in to the "Search generative AI" setting in Search Console. (Google Search Central)
Analytics
GA4 groups visits from AI tools into a default channel called AI Assistant, covering ChatGPT, Gemini, DeepSeek, Copilot and Grok. It excludes Google's AI Overviews and AI Mode, which show up as organic search. (Google Analytics Help) Watch landing pages and conversions for that channel, and expect a lot of homepage arrivals.
Conversion rates from AI visits can be high, though one company's numbers won't predict yours. Ahrefs reports that 6% of its AI search visitors sign up for a free account, four times its organic search conversion rate. (Ahrefs)
Step 5: Score accuracy and fix the source
Each month, read the answers marked "partly wrong" or "wrong" and sort the errors:
Out-of-date facts, such as old pricing, a retired plan or a renamed feature.
Missing facts, such as an integration the answer says you lack.
Borrowed facts, where a competitor's feature or limitation is attached to you.
Category errors, where the answer puts you in the wrong category or segment.
Then trace each one to its source. Where the engine shows citations, the error often comes from an old press release, a directory listing or a third-party comparison page. Fix your own page first, putting the correct fact in plain words near the claim it qualifies, then ask third parties to update theirs.
Don't respond by publishing a page for every prompt you track. Google's AI optimization guide warns that content created for each query variation "primarily to manipulate rankings or generative AI responses" qualifies as scaled content abuse. (Google Search Central)
Step 6: Connect visibility to pipeline
Visibility is a leading indicator. It tells you whether you're in the answers that shape a shortlist. It doesn't tell you about revenue, and a tracking tool can't either. Google's guide notes that "No third-party tool has access to our internal ranking or AI systems." (Google Search Central) Tools sample answers, just as you do by hand.
To get from visibility to pipeline:
Add AI assistants (ChatGPT, Perplexity, Gemini, Claude) as options in the "How did you hear about us?" field on demo and signup forms, with a free-text box.
Pass first-touch source and landing page into the CRM.
Compare visibility trends for a prompt group with demo volume from the matching pages and self-reported AI answers, over quarters rather than weeks.
Kevin Indig made the point on X in September 2026: "referral traffic is not the ideal way to judge the impact of AEO. self-reported attribution is much more useful and that has a long lead time." (Kevin Indig on X)
The B2B SEO pipeline attribution guide covers the full four-layer setup, including the CRM side, a sample monthly report and how to present it to a CFO.
A monthly AI visibility report on one page
Keep the report short enough that someone reads it:
Share of voice by prompt group (shortlist, alternatives, integration, problem, pricing), against your top three competitors, with the three-month trend.
Citation rate, and which of your URLs get cited most.
Accuracy: the count of wrong or partly wrong answers, the top errors and their status.
Google AI impressions from Search Console, by page.
AI Assistant sessions and conversions from analytics.
Self-reported AI mentions on demo forms, and the pipeline attached to them.
Known model or product changes during the month.
Lines 1 to 3 come from your prompt tracking, lines 4 and 5 from Google, and line 6 from the CRM. Keep them separate on the page, so nobody mistakes a sampled estimate for a measured number.
Tracking by hand or with a tool
A spreadsheet and a few hours a month is enough to start with 30 to 60 prompts across four or five engines. It also forces someone to read the answers, which is how accuracy problems get found.
Tools help once the prompt set grows or you need daily sampling. Ahrefs Brand Radar, Semrush's AI Visibility Toolkit and Similarweb all sell AI visibility tracking, and plenty of newer tools do too. When you compare them, ask how many runs they take per prompt, from which locations, whether they log the full answer text, and whether you can export the raw data. If a vendor can't explain how it samples, treat its scores with care.
If you'd rather hand this to an agency, the lists of AEO agencies for B2B SaaS, generative engine optimization agencies and AEO agencies for agentic AI companies compare who can show citation reporting. Ask any agency to show a real report, and be wary of anyone who promises a citation. Miraa's own AEO page says it doesn't promise one. (Miraa)
Common measurement mistakes
Tracking branded prompts only. "What is [your brand]?" will nearly always mention you and tells you nothing about shortlists.
Changing the prompt set every month, which wipes out the baseline.
Recording one run per prompt and reporting it as fact.
Counting mentions without checking whether the description is correct.
Treating a rise in AI referral traffic as the whole story, when much AI influence arrives as direct or branded search.
Reporting tool scores to leadership as if they were measured traffic.
Start with a hand-run baseline
AI search visibility can be measured well enough to make decisions, as long as you're honest about what each number is. Prompt tracking is a sample, Search Console is first-party but Google-only, and pipeline comes from the CRM and from buyers telling you where they heard about you.
This month, build a 40-prompt set from real sales conversations and run it by hand three times per engine. Read the answers. Fix the accuracy errors you find before you chase any visibility score; a wrong description that's cited often is worse than no citation. Buy a tracking tool only once you know which prompts matter and have run them by hand at least twice.
For how AEO and SEO metrics differ, see AEO vs SEO.