Know exactly how often AI recommends you — and why.
Search Console tells you where you rank on Google. Nothing tells you, out of the box, how often ChatGPT, Gemini, Perplexity, Copilot or Google's AI Overviews name your brand, cite your pages or describe you accurately. Our AI visibility tracking service measures it properly — across engines, prompts, locations and repeated runs — and ties it to real traffic.
the same brand list (SparkToro)
prompt set
& reporting
There is no "rank #3" in an AI answer. There's a probability.
Ask an AI assistant the same question twice and you'll often get two different lists. That isn't a bug in your tracking — it's how generative models work. Any report built on a single screenshot is measuring luck.
The good news from the same research: while individual answers are noisy, how often a brand appears across many runs is far more consistent. In the study's headphone prompts, well-known brands appeared in a majority of responses even though the exact list changed every time. That's the metric that matters — your share of the consideration set.
So we measure visibility the way pollsters measure opinion: a fixed, representative question set, repeated sampling, and trends reported with their margin of noise — not one-off anecdotes.
It also means tracking has to cover several engines at once. A brand can be a fixture in Perplexity's cited sources and almost absent from Gemini, or well described by ChatGPT but misquoted by Copilot because Bing still ranks an outdated directory page. Each engine retrieves from a different index, weights sources differently and updates on its own schedule. Looking at only one of them — usually ChatGPT, because it's the one leadership asks about — hides most of the picture.
The questions this service answers
- For the prompts that drive revenue, how often are we named — and by which engines?
- Who wins when we don't, and which sources do engines cite for them?
- Is what AI says about our pricing, features and reputation correct?
- How much traffic and how many leads actually come from AI platforms?
Six numbers that describe your AI visibility.
Each one is defined precisely and calculated the same way every month, so trends are real and comparable across engines.
Mention rate
Share of answers that name your brand at least once.
answers mentioning you ÷ total answers
Share of voice
Your mentions as a share of all brand mentions in your competitive set — the closest thing to "market share" inside AI answers.
your mentions ÷ all tracked-brand mentions
Average position
Where you appear when you're named, averaged across runs. Useful only alongside mention rate — first in 10% of answers is worse than third in 70%.
mean list position when present
Citation share
Share of answers that link to your domain as a source — and which of your URLs earn those links.
answers citing your domain ÷ total answers
Sentiment
Is the description positive, neutral, negative or qualified ("good but expensive")? Classified per answer, with human review of edge cases.
positive / neutral / negative split
Accuracy
Fact checks against your source of truth: pricing, locations, features, founding year, leadership. Every error is logged with the likely source.
correct claims ÷ checkable claims
One view of every engine, prompt and competitor.
Below is a mock-up of the monthly dashboard we deliver. All figures are sample data for a fictional B2B software brand — they illustrate the format, not a real result.
Share of voice · all engines
Sample data · 120 prompts × 5 runs × 5 engines
| Engine (sample) | Mention rate | Citation share | Accuracy |
|---|---|---|---|
| ChatGPT | 38% | 12% | 96% |
| Gemini | 31% | 9% | 94% |
| Perplexity | 44% | 27% | 98% |
| Copilot | 29% | 15% | 81% |
| AI Overviews | 22% | 18% | 95% |
Sample data. In this fictional example the Copilot accuracy dip would trigger an investigation — often an outdated directory listing that Bing still ranks.
Prompts where the brand appeared in at least 3 of 5 runs — our "reliably visible" threshold.
Drop in mention rate on "pricing" prompts after a competitor's comparison page launched.
Third-party pages cited most often for the brand's category — the PR target list for next month.
Good tracking starts with the right questions.
Your prompt set is the sample frame for everything else. Get it wrong and every chart afterwards is precise and useless.
- Sourced from real demand. Sales-call notes, site search, support tickets, Search Console queries and People Also Ask — rewritten the way people talk to assistants.
- Balanced across the funnel. Category ("what is…"), shortlist ("best … for …"), comparison ("X vs Y"), and decision ("is X worth it", "X pricing").
- Mostly unbranded. Branded prompts check accuracy; unbranded prompts measure whether you're discovered at all.
- Tagged for analysis. Each prompt carries stage, product line, persona and market so you can slice results.
- Frozen, then versioned. We hold the core set stable for trend lines and add new prompts as a separate cohort.
- Includes follow-ups. Real conversations continue — "which of those is cheapest?", "any with Indian support?" — so we track short multi-turn sequences for key journeys, not just opening questions.
See more templates in our AI search prompt library.
Get a free AI visibility snapshot.
Send us your category and three competitors. We'll run 20 buyer prompts across five engines, multiple times each, and show your mention rate, share of voice and any inaccurate claims.
Controlling the variables that distort AI tracking.
AI answers change with who's asking, where, when and how. Here's how we keep the measurement honest.
| Variable | Why it matters | How we handle it |
|---|---|---|
| Non-determinism | The same prompt returns different brands from run to run. | Each prompt runs multiple times per cycle (typically 3–10); we report rates and flag changes smaller than normal run-to-run noise. |
| Logged-in vs logged-out | Memory, personalisation and plan tier can change answers. | A clean logged-out baseline plus fresh logged-in accounts with memory off; never staff accounts. |
| Location & language | Local and pricing answers vary by country and city. | Runs from each target market and language you sell in, including Hindi or regional languages where relevant. |
| Model & mode | Default models, search on/off and AI Mode vs Overviews behave differently. | We log model and mode for every run, and separate web-search answers from model-only answers. |
| Timing | Fresh news and index updates move answers week to week. | Fixed weekly schedule, with event-based re-runs after launches or PR wins. |
| Prompt wording | Small phrasing changes can swing results. | Two to three variants per intent, analysed as one group. |
A note on tools
Several third-party platforms now automate AI prompt tracking, and we use them where they fit. But tool numbers differ because methods differ — runs per prompt, API vs. interface, location, logged-in state. We document our method so your numbers mean something, and compare tools in our AI SEO tools guide.
Tracking the traffic AI sends you.
Mentions are leading indicators. Visits, leads and revenue are the outcome. We connect the two with analytics, first-party search-engine reports and server logs.
Separate AI traffic from "referral" noise
In May 2026 Google Analytics introduced an AI Assistant channel in the default channel group, so chatbot referrals such as ChatGPT, Gemini and Claude are no longer lumped into generic referrals. We still build a custom channel group and explorations, because the default list may not catch every assistant and you'll want history from before the change.
ChatGPT also appends utm_source=chatgpt.com to links it shows, according to OpenAI's publisher FAQ, which helps attribution even when the referrer is stripped.
Sources: Google Analytics Help — What's new · OpenAI Publishers & Developers FAQ
| Platform | Typical referrer / source |
|---|---|
| ChatGPT | chatgpt.com · utm_source=chatgpt.com |
| Perplexity | perplexity.ai |
| Gemini | gemini.google.com |
| Microsoft Copilot | copilot.microsoft.com |
| Google AI Overviews / AI Mode | Arrives as Google organic — not separable in GA4 |
Referrers can be stripped by apps and privacy settings, so GA4 undercounts AI visits. Treat these as a floor.
Google Search Console
Google says traffic from AI Overviews and AI Mode is included in the Performance report under the "Web" search type. Google has also introduced a Generative AI performance report showing impressions where links to your site appeared in AI Overviews and AI Mode, by page, country and device; Google's help page says it rolled out globally as of 31 August 2026, though not every site has data yet.
Sources: Google Search Central · Search Console Help
Bing Webmaster Tools
Launched in public preview in February 2026, AI Performance shows total citations, average cited pages, grounding queries and page-level citation activity across Microsoft Copilot, AI summaries in Bing and select partners. It's the first first-party citation data from a major AI assistant — we add it to every report.
Source: Bing Webmaster Blog, Feb 2026
See which AI bots read which pages
Server or CDN logs show what analytics can't: AI crawlers fetching your content. Spikes in user-initiated fetches on a page often mean it's being used to answer live questions. We verify bots against published IP ranges to filter spoofed user agents.
| User agent | Purpose (per vendor docs) |
|---|---|
| OAI-SearchBot | Surfaces sites in ChatGPT search |
| ChatGPT-User | User-initiated fetches in ChatGPT |
| GPTBot | Collects data for model training |
| PerplexityBot | Indexes for Perplexity search |
| Perplexity-User | Fetches pages when users ask questions |
| Bingbot / Googlebot | Feed Copilot and Google's AI features |
Sources: OpenAI crawler docs · Perplexity crawler docs
Every pattern in the data points to a fix.
A tracking report is only worth paying for if it changes what your team does next month. These are the patterns we see most often, what usually causes them, and the work that addresses them.
| What the data shows | Likely cause | What we recommend |
|---|---|---|
| Named in branded prompts, absent from "best X for Y" | Weak third-party consensus — few reviews, listicles or articles put you in the category. | Earn placements on the sources engines cite; see AI brand mentions. |
| Mentioned often, cited rarely | Engines know the brand but your pages aren't the most quotable answer. | Answer-first rewrites and original data via citation optimization. |
| Strong in Perplexity & ChatGPT search, weak in model-only answers | Recent visibility not yet reflected in model training data. | Keep building consistent mentions; re-check after new model releases. |
| Weak only in Copilot | Bing indexation gaps or outdated pages ranking in Bing. | Bing Webmaster Tools, IndexNow, clean-up — see Copilot SEO. |
| Absent from AI Overviews despite good rankings | Answers buried in long pages; sub-questions from query fan-out not covered. | Passage-level structure and supporting pages — see AI Overviews optimization. |
| Recurring wrong price or feature | Old pages, directories or reviews still state outdated facts. | Update the source, publish a clear facts page, align schema and profiles. |
| Negative qualifier ("but support is slow") | Review themes or forum threads the engine summarises. | Fix the operational issue, respond publicly, encourage recent reviews. |
Visibility is a leading indicator
AI mentions often show up before traffic changes, because many people read the answer and act later — searching your brand directly or visiting from another device. We look at branded search demand and direct traffic alongside AI referrals, rather than judging success on last-click AI visits alone.
What tracking can't tell you
No tool sees every real user conversation; engines don't publish query volumes for AI prompts. Prompt tracking is a well-designed sample, not a census. We're explicit about that, and we never present sampled rates as market-wide truths.
Set up in three weeks. Reported every month.
Discover
Week 1. Interviews with sales and marketing, competitor set, markets, languages and a source-of-truth fact sheet for accuracy checks.
Design
Week 1–2. Build and tag the prompt set (typically 50–200 prompts), choose engines, locations, run counts and logged-in/out mix.
Baseline
Week 2–3. First full measurement cycle, GA4 AI channel set-up, Search Console and Bing Webmaster Tools connected, log analysis configured.
Report & act
Monthly. Trends, wins, losses, accuracy errors and a short prioritised action list for content, PR and technical teams.
Weekly
Automated runs of the full prompt set. Alerts for sudden drops, new competitors in answers, or new inaccurate claims about your brand.
Monthly
Dashboard and written report: six core metrics by engine, stage and market; AI referral traffic and conversions; top cited sources; next actions.
Quarterly
Strategy review: prompt-set refresh, competitor re-benchmark, and ROI view linking visibility changes to leads and revenue.
What's included.
- Prompt set — documented, tagged and versioned, owned by you.
- Live dashboard — mention rate, share of voice, average position, citation share, sentiment and accuracy by engine, market and funnel stage.
- Accuracy log — every incorrect claim, the engine, the likely source and the fix status.
- Source map — the third-party URLs cited most in your category, refreshed monthly.
- AI traffic reporting — GA4 AI channel and custom groups, conversions and assisted revenue from AI platforms.
- First-party AI reports — Search Console generative AI data and Bing AI Performance citations folded into one view.
- Bot log analysis — which AI crawlers fetch which pages, plus robots.txt recommendations.
- Monthly call with your analyst and a clear, prioritised action list.
Measurement built to drive action, not decorate a slide.
Analytics roots
We've built search and paid-media reporting since 2008 and work as a Google Ads Certified Partner — attribution is part of our daily work.
Statistically honest
Repeated runs, documented methods and noise thresholds. We'll tell you when a change isn't real yet.
Tracking that leads to fixes
The same team runs citation optimization, brand-mention building and engine work like ChatGPT SEO.
Multi-market
Experience across 98+ industries and many countries, with location- and language-specific runs where you sell.
Want to see measurement in context? Read our AI search experiments and GEO statistics.
AI visibility tracking, answered.
What is AI visibility tracking?
AI visibility tracking measures how often, where and how accurately AI assistants and answer engines — such as ChatGPT, Gemini, Perplexity, Microsoft Copilot and Google AI Overviews/AI Mode — mention and cite your brand for a defined set of buyer prompts. It reports metrics like mention rate, share of voice, average position, citation share, sentiment and accuracy, alongside the traffic AI platforms send to your site.
Why do AI answers change every time I ask?
Generative models produce answers probabilistically, and engines that search the web may retrieve different sources each time. Research by SparkToro and Gumshoe found identical prompts returned the same brand list less than 1% of the time. That's why we run each prompt several times and report rates rather than single answers.
How many prompts should we track?
Most businesses start with 50–200 prompts covering their main products, markets and funnel stages. Fewer prompts with repeated runs beat a huge list run once. We expand the set as your offering and markets grow.
Can I see AI Overviews traffic in Google Analytics?
Not separately. Clicks from AI Overviews and AI Mode arrive from Google and are counted as organic search. Google includes them in Search Console's Performance report under the Web search type and has introduced a Generative AI performance report showing impressions in those features where data is available.
Does GA4 track ChatGPT and Perplexity traffic?
Yes, when a referrer or UTM tag is passed. Google Analytics added an AI Assistant channel to the default channel group in May 2026, and ChatGPT appends utm_source=chatgpt.com to links. Some visits still lose their referrer, so we treat GA4 figures as a minimum and add custom channel rules.
Do you track logged-in or logged-out answers?
Both, separately. Logged-out runs give a clean baseline; fresh logged-in accounts with memory off reflect how many real users experience the product. We never use personal or staff accounts, whose history would bias results.
Which tools do you use?
A combination of third-party AI tracking platforms, our own scripted runs, GA4, Google Search Console, Bing Webmaster Tools and server-log analysis. We choose per client and document the method so numbers are comparable over time.
How quickly will I see results?
Your baseline is ready within about three weeks. Tracking itself doesn't change visibility — it shows where to act. Clients who pair tracking with optimisation work typically see measurable movement on web-search-driven engines within a few months.
Can we start with a one-off audit?
Yes. A one-off AI visibility audit gives you a baseline across engines, a competitor comparison, an accuracy check and a prioritised action list. Many clients then continue with monthly tracking.
Keep exploring generative engine optimization
Every AI platform rewards slightly different signals. Dive into the playbook for each one — or let our GEO team run all of them for you.
Find out what ChatGPT, Gemini & Google AI say about you today.
Tell us your website and top competitors. Within 3 business days a GEO strategist sends you a personal AI visibility report — no generic PDF, no sales script.
- Share-of-voice snapshot across ChatGPT, Gemini, Perplexity, Copilot and Google AI Overviews
- Prompt-level gaps where competitors are named and you aren't
- Citation sources each AI engine leans on in your category
- A prioritised 90-day GEO plan you can run in-house or with us