What a baseline is for
Most teams start with screenshots: somebody asked an assistant about the category and did not like the answer. A baseline replaces that with a record you can compare against next month.
It should end in three short lists: where you are strong, where you are missing, and what is being said against you. Expect it to be less flattering than the screenshots, and expect at least one surprise in each list.
Presence: are you named at all?
Start with the category and comparison prompts. For each claim and each engine, look at two things together: how often you are named, and where in the answer you appear. Named first some of the time is a different situation from named third every time.
Then look claim by claim, starting with the ones you prioritised. An overall figure hides the pattern that matters, which is usually strong on one claim and absent on another.
Finally, test the category word itself. Buyers describe the same kind of product in different ways, and you can be well represented under one label and missing under another.
Sources: what the engines are reading
For the category prompts, collect every source the engines cited. Leave branded prompts out of this step: they mostly cite your own site and will make it look more influential than it is.
Group what is left by kind: review sites, forums and communities, comparison and list articles, documentation, news, competitors' pages, your pages. The pages that come up again and again are the shortlist your work has to reach. Most of them will not be yours.
What the answers say about you
Read the answers that mention you, not just the fact that they do. Sort what they say by claim. Three things usually come out of it:
- Objections that recur: too expensive, hard to set up, weak support. These need an answer, not a denial.
- Facts that are wrong or out of date: old pricing, a discontinued plan, a feature you never had.
- Strengths you under-sell: things engines praise that your own site barely mentions.
When engines disagree, that is useful
Compare the engines claim by claim. Where they all agree about you, the picture is settled, and it will be slow to move. Where they split, the claim is still open, and that is usually where effort pays off first.
Resist optimising for one engine. They weigh sources differently and change often; work that strengthens the claim in the sources they share lasts longer than work aimed at one model's habits.
Crawlers and visits
Two more views complete the picture, and both are partial. Your server logs show which AI agents fetch which pages, and whether a fetch is a model learning your site or quoting it in a live answer. A page that is never fetched cannot be quoted. Your analytics show visits referred by assistants, which undercount influence but confirm that answers do send people.
Hand the findings to the people who own them
Almost none of the fixes sit with one team. Route each finding to its owner:
| Finding | Usually owned by |
|---|---|
| Wrong or outdated facts about the product | Product marketing and documentation |
| Recurring complaints in reviews | Customer success and support |
| Missing from comparison and list articles | PR and partnerships |
| Own pages that do not state the claim clearly | Content and SEO |
| A category label buyers do not use | Brand and product marketing |
Frame every request the same way: where it stands now, what it should say, why their team is the one who can change it, and the specific thing you are asking for.
Key takeaways
- A baseline ends in three lists: strengths, gaps and objections.
- Read how often and where you are named together, claim by claim, and under more than one category label.
- Analyse sources from category prompts only; the recurring pages are your shortlist.
- Split verdicts between engines mark the claims most open to change.
- Route each finding to the team that owns the source.