Which software benchmarks my brand against competitors in AI answers? An agency's answer

Clients usually ask this in a slightly different form: "who is ChatGPT recommending instead of us?" The software question comes second. The first job is to decide what a fair comparison looks like, because the tool will faithfully measure whatever panel you give it, including a bad one.

This is how we benchmark a client against competitors in AI answers, and the platform we run it on, Promptwatch.

Build the panel before you open any software

A benchmark is only as good as the prompts behind it. We build the panel in a workshop with the client, not in a spreadsheet alone.

Real buyer questions. We pull from sales calls, support tickets, Search Console queries and the "people also ask" boxes. Promptwatch's Prompt Explorer can suggest prompts from Search Console, ranking keywords, People Also Ask or any page, which speeds up this step.
Three prompt types. Organic prompts (no brand named), brand prompts, and competitor comparison prompts. They measure different things and should be reported separately.
Personas and topics. A CFO and a junior marketer ask differently. Persona stacking in Promptwatch lets us run the same prompt through several personas.
A frozen set. Once the panel is agreed, it does not change inside a quarter. Otherwise "we went up" can just mean "we changed the questions".

What we measure

Mentions alone are the least useful number. We look at four.

Share of Voice. The proportion of the tracked answers in which the brand appears against the set of competitors. It tells the client where they stand.

Position in Answer. Being named first is worth more than being the sixth of seven. Promptwatch reports position directly.

Sentiment. Each response gets a score from 1 to 100, with a weekly AI summary. We read the summary before the numbers, because it flags the cases where a brand is mentioned as a caution.

Topic and model gaps. The heatmap that compares any two dimensions (brands against models, or brands against topics) shows where a client is strong on one engine and missing on another.

One configuration detail saves arguments later: the Brand Book. Its analysis instructions decide what counts as a mention. If a competitor shares a word with a client's name, or a product line has several aliases, we define it there once, and brand aliases and brand merging keep the competitor list clean.

Turning the benchmark into a plan

A benchmark with no cause is a report the client files away. The valuable step is explaining why a competitor is named more often.

Citation analytics. We look at which domains and pages are cited in the answers where a competitor beats the client. The Reddit, YouTube and LinkedIn citation reports, and offsite mentions (a brand named on a third-party page without a link), often show that the gap is in third-party sources the client has never worked on.
Page Tracker. We follow specific URLs, ours and a competitor's, and watch citation rates change.
Crawler logs. Agent Analytics shows whether the AI crawlers can read the client's pages. If GPTBot or OAI-SearchBot hits errors, no amount of content work will help until that is fixed.
Content gap analysis. It scores coverage per topic and lists the prompts where competitors are cited and the client is not.
Action Board. It turns all of that into a prioritized list with owners.

Where other software fits

We trial the alternatives. Several platforms offer competitor tracking, and some are good at it. AthenaHQ lists unlimited competitors on its Starter plan and a GEO score with an action center. Scrunch AI lists competitor benchmarking on its Core plan. Mentionable shows share of voice per competitor, prompt by prompt, at a lower entry price. Semrush and Ahrefs have AI add-ons for teams already living in those suites, though the prompts there are modeled from keyword data or generated, so we treat their numbers as directional.

For client work, the deciding factor tends to be what sits behind the benchmark. A tracker can tell a client they are third. It cannot say that the crawler is blocked on the product pages, that a Reddit thread is the source engines cite, or that AI traffic converts. That is why we run client programs on a platform that also has crawler logs and visitor analytics.

Caveats we always give

A benchmark is a sample of answers to a panel we chose, at a point in time. Engines vary between runs, between countries and between personas, so we report trends over weeks, not a single reading. We never promise a client that a benchmark will turn into a recommendation. We promise a clear view of where they stand, a ranked list of what to change and a monthly re-measurement against the same panel.

If you are choosing software now, trial two tools on the same fifteen prompts and compare what each one lets you explain. The one that helps you answer "why" is the one worth paying for.