A baseline tells you what AI says about you. A benchmark tells you what it says about you compared with the companies you lose deals to. The second is what changes decisions: being named in a third of answers sounds fine until you find out your main competitor is named in two thirds.
This is the method we use. It builds on our baseline guide, so if you haven't written your question set yet, start there.
Step 1: choose two or three competitors
Pick the companies you actually lose to, not the biggest names in your market. Ask sales who comes up in deals, and who prospects mention when they say "we're also looking at". Two or three is enough. More makes the benchmark slow to run and hard to read.
Step 2: fix the question set and the settings
Use the same twenty or so buyer questions for every competitor, in the same words. Most of them won't mention any brand: "what's the best X for Y?" is where recommendations happen. Add a couple of direct comparisons ("X vs Y") for each competitor.
Then fix the settings: which AI tools, which versions, signed in or out, web search on or off, and which country. Write them down. If any of these change between runs, the comparison breaks.
Step 3: record every answer the same way
One row per question and model. For each, record:
| Column | What to write |
|---|---|
| Named | Which of you and your competitors appear |
| Order | Who's mentioned first, second, third |
| Framing | A few words on how each is described |
| Sources | The sources the answer cites or links |
| Errors | Anything inaccurate about any of you |
Keep the full text of each answer in a second tab. You'll want it when you're working out why something changed.
Step 4: calculate share of answer
For each brand, divide the number of answers it's named in by the total number of answers. That's its share of answer. Do the same for first mentions, which matter more than any mention.
A simple summary table is enough:
| Named in | First mention | |
|---|---|---|
| You | 11 of 40 | 3 of 40 |
| Competitor A | 26 of 40 | 14 of 40 |
| Competitor B | 18 of 40 | 9 of 40 |
An example layout, not real data.
Step 5: find the source gap
This is where the benchmark earns its keep. List every source cited in answers that name a competitor but not you. Group them: forums and communities, review platforms, press, reference pages, directories, competitors' own sites.
The result is usually a short, specific list. A comparison thread where a competitor is recommended and you aren't mentioned. A review platform where they have far more recent reviews. A trade article that lists them and not you. Each one is a task.
Step 6: turn the gap into a plan
Rank the gaps by two things: how often the source is cited, and how hard it is to change. A review platform cited in half the answers, where you simply haven't asked customers for reviews, comes first. A major publication that would need a news story to mention you comes later.
Step 7: repeat on a schedule
Run the same benchmark monthly, or quarterly if your market moves slowly. Same questions, same competitors, same settings. Look at trends, not single runs. Answers vary between runs, so one month's jump might be noise. Three months in the same direction is a signal.
Common mistakes
- Benchmarking against the market leader you never meet in deals. Interesting, but not useful.
- Changing the questions halfway through. You lose the trend.
- Only counting mentions. A competitor named with "people complain about their support" is not ahead of you.
- Skipping the sources. Without them you know the score but not what to do.
What this doesn't mean
A benchmark shows where you stand. It doesn't move anything on its own, and a lower share than a competitor isn't a verdict on your product. It usually means they've been talked about in more of the places models read.
If you'd like this done for you, with every source traced and a ranked plan at the end, that's our visibility audit: five models, you and two competitors, one week. And for what to measure after that, see the GEO KPIs we use.