A SaaS marketing team can usually tell you their position three, their domain rating, and their click-through rate on branded search. Ask the same team how often ChatGPT recommends them against a named competitor, and most go quiet. Not because the question doesn’t matter: because nobody is running it on a schedule, logging what comes back, or comparing this month to last month.
That gap is easy to miss because checking manually feels like enough. Someone opens ChatGPT, types a buyer-style question, sees the brand mentioned, and reports “we’re showing up in AI search.” One favorable answer is not a measurement system, it is a single sample, and Semrush’s 2026 AI Visibility Index, built from 126 million US AI search prompts run across ChatGPT, Gemini, Google AI Mode, and AI Overviews between January and April 2026, found that only 36 global brands held top-100 visibility on all four platforms in every month studied. Visibility on one engine, on one day, tells you almost nothing about visibility on the other three.
This post is a practical framework for building that measurement system yourself: a fixed prompt panel, a sampling discipline, a log schema, and the metrics worth trending. It also covers where a DIY tracker stops being worth the engineering time and a platform like Ahrefs Brand Radar or Profound takes over. If you want the cadence for running whatever tracker you end up with, our AI Search Monitoring cadence framework covers the operating rhythm this sits inside of.
What AI visibility tracking actually measures
Before building anything, it helps to use the same vocabulary the commercial platforms use, since “we’re visible in AI search” hides at least four different, non-interchangeable signals. Ahrefs documents its Brand Radar definitions this way, and they’re worth adopting wholesale rather than inventing your own terms:
- Mention: your brand name appears as a string in the AI’s answer text, linked or not.
- Citation: the answer includes a linked URL to your domain, whether or not your brand name is actually said.
- Found, but not cited: the answer is clearly shaped by your content, but no link appears, so the brand gets the influence with none of the referral traffic.
- Share of voice: how often your brand is mentioned or cited relative to every other brand that showed up across the same prompt panel.
Mentions and citations diverge more than most teams expect. Semrush’s 2026 Ghost Citations study, based on 3,981 domain appearances across 115 prompts in 14 countries, found that 61.7% of citations across ChatGPT, Google AI Overviews, Gemini, and Google AI Mode never produced a matching brand mention in the answer text itself. The same research found that ChatGPT cites roughly 15 sources per response on average, leaning heavily on Reddit and Wikipedia, while Gemini cites closer to 3 sources from a narrower pool, and that on Gemini specifically the overlap between which brands get mentioned and which domains get cited can run as low as 30%. A tracker that only checks “did they say our name” misses a different, often larger signal: whether your page is quietly feeding the answer without getting credit.
Why one prompt check tells you almost nothing
The instinct to run a prompt once and treat the answer as a verdict is understandable, and it is also the most common mistake in this kind of monitoring. Large language model output is not fully deterministic, even when a caller sets temperature to zero. ML6’s analysis of OpenAI’s API traces this to floating point math and request batching on the serving side, meaning the same prompt sent twice, under identical settings, can still return different text and different cited sources.
Layer platform divergence on top of that variance and a single check becomes close to meaningless. A prompt that surfaces your brand in ChatGPT this morning may not surface it in Gemini this afternoon, and neither result predicts what Google AI Mode will say about the same topic next week. Tracking AI visibility credibly means treating every individual response as one noisy data point and building toward a trend, not a screenshot.
The SIGNAL framework for building a tracker
SIGNAL is a six-step structure for running this as a repeatable system instead of a one-off check. Each letter is a decision point; skipping any of them is usually where DIY trackers fall apart within a month.
The log schema and what to build it on
The part of this that actually takes engineering time is not calling an AI model, it is deciding what to record and sticking to the schema once you’ve picked it. A log table that works for this needs the following fields at minimum:
| Field | Purpose |
|---|---|
| run_id | Unique ID per API call, so repeated samples of the same prompt don’t overwrite each other. |
| prompt_id | Stable identifier tied to the fixed wording, not the topic, so a phrasing change is visible in the data. |
| engine | Which platform and model version answered. |
| timestamp | When the call ran, to build weekly and monthly aggregates. |
| result_type | mention / citation / found_not_cited / absent, graded consistently per the definitions above. |
| cited_url | The exact linked page, if any, so you can see which page is actually earning citations. |
| position | Where in the answer the mention or citation landed: first source, buried in a list, etc. |
| raw_response | The full text, stored for later re-grading if your taxonomy changes. |
One honest limitation worth stating up front: querying a platform’s official API is not identical to what a human sees in the consumer product. Browsing-enabled answers, personalization, and locale can all differ between the API and the ChatGPT or Gemini app a buyer actually uses, and automating queries against a consumer chat interface instead of an official API usually runs against that platform’s terms of service. Build against the APIs that exist for this purpose, treat the gap between API and consumer-app behavior as a known blind spot, and say so in whatever you report internally rather than presenting API results as a perfect proxy for the live product.
Here is what the output looks like once a few cycles have run. The numbers below are invented for illustration and are not a real client’s data.
| Engine | Mention Rate | Citation Rate | Top Cited Page |
|---|---|---|---|
| ChatGPT | 44% | 27% | /blog/pricing-guide |
| Gemini | 29% | 12% | /blog/integration-guide |
| AI Overviews | 41% | 24% | /blog/comparison-page |
DIY tracker vs. buying a platform
A DIY tracker is worth building when the panel is small and the question is specific. It stops being worth maintaining once the scope grows. Here’s a practical comparison:
| Approach | Best for | Platforms covered | Tradeoff |
|---|---|---|---|
| DIY, SIGNAL-style | 50 to 150 prompts, 2 to 3 engines, one brand | Whatever has an official API you build against | Free beyond API costs, but you own every bug, outage, and schema change |
| Ahrefs Brand Radar | Teams already on Ahrefs wanting PAA-sourced prompts at scale | ChatGPT, Gemini, Perplexity, Copilot, AI Overviews, AI Mode | Prompts are drawn from search demand, not written by you, so niche questions may be excluded |
| Profound | Teams wanting sentiment and context analysis alongside raw visibility | Major consumer AI platforms | Premium pricing tier; see our full comparison of Profound alternatives |
| Semrush AI Toolkit | Teams wanting AI visibility next to existing SEO data | ChatGPT, Gemini, AI Mode, AI Overviews | Industry benchmarks are useful, but per-prompt granularity is coarser than a custom panel |
The honest line to draw: if you can name the 50 to 150 questions that actually matter to your pipeline and you have engineering time to maintain a scheduled job, build it. If you need coverage across dozens of markets, don’t have the engineering bandwidth to babysit API changes, or want sentiment and competitive benchmarking built in, buying a platform is the faster path to a defensible answer.
Either way, the output is only useful if someone owns reading it on a schedule. That’s the layer our AI Search Monitoring cadence framework covers once the tracker itself, DIY or purchased, is producing data.
Where this fits if you’d rather not build it
Teams that want the SIGNAL framework built and wired into their own analytics stack, rather than run as a side project between other priorities, are exactly what our Customized Analytics MCP and API work is for: we build the prompt panel, the sampling job, and the log schema against your actual buyer language, then hand you the workflow outright. If you want a lighter starting point first, a GEO audit gives you a current citation snapshot in five days before you decide whether a full tracker is worth building. For a broader conversation about what a tracker should actually be answering for your pipeline, you can book time with our team.
Frequently asked questions
What is an AI visibility tracker?
An AI visibility tracker is a system, bought or built, that repeatedly runs a fixed set of prompts through AI engines like ChatGPT, Gemini, Perplexity, and Google AI Overviews, then logs whether your brand was mentioned, cited with a link, or missing entirely. It turns an anecdotal screenshot into a dataset you can trend over time.
What is the difference between an AI mention and an AI citation?
A mention is your brand name appearing in the AI’s answer text. A citation is a linked URL to your site in that answer or its source list. The two do not move together: Semrush’s 2026 Ghost Citations study found that 61.7% of citations across ChatGPT, Gemini, AI Overviews, and AI Mode never produced a matching brand mention in the answer text.
How many times should you run each prompt to get a reliable reading?
Run each prompt multiple times per sampling cycle rather than once. Large language models are not fully deterministic even at a temperature of zero, since floating point and batching behavior on the serving side introduce variation, so a single response is one sample from a distribution, not a settled answer.
Can a small marketing team build an AI visibility tracker instead of buying one?
Yes, for a narrow panel of 50 to 150 prompts across two or three engines, a small team can build and run a tracker using official APIs, a structured log table, and a weekly cadence. It gets expensive to maintain once you need dozens of platforms, hundreds of markets, or audit-grade historical coverage, which is the point where a platform like Ahrefs Brand Radar or Profound starts to pay for itself.
How often should an AI visibility tracker actually run?
Weekly is the minimum cadence that reliably separates a real citation shift from ordinary model variance. Daily runs on a small prompt panel mostly report noise, since any single engine’s answer to any single prompt can change from one call to the next regardless of anything happening on your website.
Share this article
Ready to audit your organic growth opportunity?
$2,500 flat. 5 business days. Six deliverables tied to pipeline , not rankings. No retainer required.
Get the Organic Growth Audit →