Understand how AI evaluates your category
When a buyer asks an assistant which tool to use, an answer gets composed from sources you did not pick and competitors you did not invite. TrackGeo asks those same questions on a schedule and records what comes back — which prompts you appear in, who is recommended instead, and what the model read to decide.
The evaluation now happens before the visit
A buyer describes their problem, gets a shortlist of three tools, and only then starts clicking. By the time they reach your site the comparison is over — and your analytics never saw the part that decided it.
You are described, not just listed
Assistants do not return ten links. They characterise you in a sentence or two, and that sentence is built from whatever sources the model trusts.
The criteria are chosen for you
The model decides which dimensions matter in a comparison. If it picks criteria you lose on, you lose the answer regardless of your product.
Most of it leaves no referrer
Buyers arrive later, by name or direct. The channel that shaped the shortlist is the one with the least instrumentation.
The questions that decide a shortlist
TrackGeo classifies every prompt in your portfolio into one of 13 buyer intents. These eight are the ones that decide software categories — each scored for buyer value, tracked per engine, and marked as a win, a loss, or a prompt a competitor owns.
Category discovery
The buyer does not know your category by name yet. They describe a problem, and the model decides which kind of tool solves it.
Comparison
You against a named rival. The model picks the criteria, and the criteria usually decide the winner.
Alternative
Someone asks for alternatives to a competitor. Appearing here is how you enter a deal you were never invited to.
Competitor replacement
Active switching intent — the buyer already wants out. These prompts convert, and they are the ones worth losing sleep over.
Integration
Whether you work with the stack they already run. A wrong answer here removes you before a human ever reads your docs.
Pricing and buying
What you cost and how you are sold. Models answer this from whatever they can find, which is not always your pricing page.
Use case
Whether you fit a specific job. Narrow fit questions are where a smaller competitor beats a better-known brand.
Enterprise
Security, scale and procurement questions that decide whether you survive a shortlist at a larger account.
Prompts are also placed on a five-stage funnel — awareness, consideration, evaluation, purchase intent and post-purchase — so a high-volume awareness prompt never outranks the evaluation prompt that actually closes deals. See how prompt portfolios work.
When the answer is not you
Knowing you are absent is not useful on its own. TrackGeo reconstructs the loss: who took the slot, what the model said about them, and which sources it leaned on to say it.
- Who was recommended instead, ranked by how often they displace you
- A summary of why they win, with their claims quoted from the answer
- The domains the engine cited, and how much they overlap with your own sources
- A share-of-answer leaderboard, and a head-to-head on visibility, citation share and prompt coverage
Prompt
What is the best alternative to [incumbent] for a mid-market team?
Recommended instead
Why they win
The answer cites a review roundup and a community thread, neither of which mentions your brand. Your own documentation is not among the sources the model reads.
An illustration of the shape of a displacement finding. It is not a customer result, and the brands are placeholders.
The engines behind the answers
Every engine here has a production adapter and is exercised by real prompt runs. We name exactly what we query, because the surface changes what the number means.
ChatGPT
OpenAI API
Perplexity
Perplexity API (sonar)
Gemini
Gemini API
Azure OpenAI
Azure OpenAI Service
Claude
Anthropic Messages API
Gemini with Search grounding
LimitedGemini API + google_search tool
TrackGeo measures answers from provider APIs — the models behind these assistants — not scrapes of their consumer apps. Read the full engine list and its caveats.
Gaps become owned work, not a slide
Findings are only useful if someone ships them. Every gap becomes an action with a type, an owner, a due date and a priority score you can interrogate.
Actions move through a status lifecycle that ends where it should: once work is implemented, the prompt is re-scanned and the action is only marked impact verified if the number actually moved.
How priority is calculated
A score out of 100, weighted from five inputs. The weights are published in-product so a prioritisation can be argued with.
- Business impact
- 35%
- Visibility opportunity
- 25%
- Competitive gap
- 20%
- Ease
- 10%
- Confidence
- 10%
What B2B SaaS teams ask first
Does TrackGeo scrape ChatGPT or the other assistant apps?
No. TrackGeo queries provider APIs — the same models that power those assistants — and asks them the buyer questions you care about. That is a repeatable, defensible measurement, but it is not a capture of the consumer app, and results can differ from what you see in a chat window.
How does TrackGeo decide which prompts matter for our category?
Every prompt is classified into one of 13 buyer intents — including comparison, alternative, competitor replacement, integration and pricing — and into one of five funnel stages from awareness to post-purchase. Each one carries a Prompt Value Score from 0 to 100, derived from its revenue relevance and conversion potential, so you can work the high-value prompts you are losing before the ones that never mattered.
Can we see why a competitor gets recommended instead of us?
Yes. For each losing prompt, TrackGeo records who was recommended instead, a summary of why they win, the competitor claims quoted from the answer itself, and the sources the engine cited to support them. Every finding links to that evidence rather than asking you to trust a score.
Do the answers change between runs?
Yes. Model outputs vary between runs, which is why TrackGeo samples prompts repeatedly instead of reading a single answer, and labels each finding with a confidence level rather than presenting one response as fact.
Which engines can you measure?
Six: ChatGPT, Perplexity, Gemini, Azure OpenAI, Claude, and Gemini with Search grounding. Engine access depends on your plan — ChatGPT and Perplexity are included on the free plan, and the full set is available on Growth.