Almost nobody asks an AI assistant which CRM is best. That question shows up in keyword tools because it is what people type into a search box, and the box has trained everyone to be terse. A chat interface does not, so people write what they actually mean.
What they write instead is long, constrained and slightly embarrassing. A team size, a tool they are stuck with, a budget they are slightly ashamed of, and a workflow that does not fit anything on the market. Those are the questions deciding shortlists, and almost nobody has published a clean answer to any of them.
Five shapes, in roughly the order they appear
Across the evaluations we see, buyer questions fall into five recognisable shapes. Each one has a different difficulty and a different owner.
| Shape | Example | Who currently wins it | Winnable? |
|---|---|---|---|
| Category | What is the best sales engagement platform | Incumbents, via review aggregators | No, not inside two years |
| Constrained | Sales engagement for a 12 person team selling into hospitals | Frequently nobody | Yes, in a quarter |
| Integration | What works with HubSpot and Gong without middleware | Whoever documented it plainly | Yes, quickly |
| Objection | Why do companies leave Outreach | Review sites and old forum threads | Partly, and it is worth trying |
| Vendor-specific | How much is Apollo and what is included | Third-party listings, often wrong | Yes, and you should already own it |
The pattern is consistent enough to plan around. The broader the question, the more the answer is decided by training data and third-party corroboration, both of which favour whoever has existed longest. The narrower the question, the more it is decided by live retrieval, which favours whoever wrote the clearest page.
Why the category question is a trap
It is the one everybody wants and the one that behaves worst. When an engine answers an unqualified category question it is drawing on dense, repeated associations built over years of the web talking about the same few companies. Your content does not enter that calculation in any meaningful way.
Adoption data supports treating this as a slow-moving structure rather than a contested space. The Stanford AI Index tracks how quickly these tools have been taken up and how concentrated attention is on a small number of systems. Concentrated attention on a small number of systems producing answers built from concentrated coverage of a small number of vendors is a compounding advantage, not a level field.
The mechanics of why some companies get named and others do not are set out in how ChatGPT picks which companies to name. The short version is that the category question is a popularity measurement wearing the costume of a recommendation.
The useful thing to know is which shapes you already win and which you do not, because the answer is rarely what people expect. Our free AI visibility check runs all five shapes against your category and returns the verbatim answers, so you can see exactly where the ground is soft.
What constrained questions actually look like
These are longer than anything a keyword tool will show you, because nobody types them into a search box. Real examples, lightly anonymised, from evaluations we have watched.
We use Salesforce but our SDRs live in Slack. What sequencing tool
won't require us to rip out either?
Sales engagement for a team of 9, under $1,200 a month total, must
work with a shared inbox rather than individual mailboxes.
We sell into NHS trusts. Which outreach tools have customers doing
UK public sector and won't fall over on procurement review?
Something that does call recording but stores audio in the EU and
doesn't need IT to approve a browser extension.Notice what these contain. A constraint that excludes most of the market, an existing tool, a budget or a compliance requirement, and an implicit objection. Notice also that no vendor page is written to answer them, because vendor pages are written to describe products rather than to resolve situations.
That gap is the opportunity and it is unusually cheap to close. One page that names the constraint in its heading, states plainly who it suits and who it does not, and gives the specific integration behaviour, will be retrieved for a family of these questions. The writing method that makes such a page extractable is in answer-first writing.
Where to find yours. Not a keyword tool. Your sales call transcripts and your inbound form free-text field. Every qualifying question a rep asks, and every awkward requirement a prospect volunteers, is a constraint somebody is typing into an assistant this week.
The vendor-specific questions you are already losing
Four questions get asked about named vendors more than anything else. Pricing, integrations, implementation time, and why customers leave.
The first three should be yours and frequently are not, because the information sits in a format a parser cannot read. A pricing table rendered by JavaScript, an integration list that is a grid of logos, an implementation timeline mentioned only in a gated PDF. When your own page cannot answer, the engine answers from a directory listing that has not been updated since 2023. The fix is in why your pricing page is the page AI reads most.
The fourth question is the uncomfortable one and it is worth running on yourself today. Ask each engine why companies stop using your product. Whatever comes back is the objection your sales team is already fighting, now stated by a neutral-sounding source at the exact moment somebody is forming a view.
The behaviour that is genuinely changing
Two shifts are worth planning around rather than reacting to.
The first is that buyers increasingly ask follow-up questions rather than accepting the first answer. An assistant that names three vendors will be asked to compare them, then asked about the cheapest, then asked about migration. You are not competing for one answer, you are competing to survive four turns. Surviving turn three requires specific, checkable facts on your own pages, because that is when retrieval gets narrow, and the vendor documentation on how assistants chain tool calls across a conversation makes clear how many separate retrievals a single exchange can involve.
The second is that some of these questions are starting to be asked by software rather than by people. Assistants that browse, fill forms and compare on a buyer’s behalf are already shipping, and the reporting on how they behave in practice has been steadily less breathless and more useful, with outlets like Wired tracking the gap between the demos and what the tools actually do. That surface is covered separately in where AI search is going for B2B SaaS.
For now the practical implication is narrow. A machine reading your pricing page has none of the patience a human has. It will not click a toggle, wait for a script or open a PDF. Everything above about plain, server-rendered facts applies twice over.
What to do with this
Take your last forty sales calls. Write down every constraint a prospect volunteered before they saw a demo. Count them. Two or three will appear repeatedly and those are your winnable questions.
Write one page per repeated constraint. Put the constraint in the heading, answer it in the first sentence, say plainly who it does not suit, and include the specific integration and pricing facts as text. Four pages like that will outperform forty posts about the category, and they take a fortnight rather than a year.
Then check quarterly whether the answers have changed, using a fixed set rather than new questions each time. The protocol is in the twenty prompt set, and for this category the modification is that most of your twenty should carry a constraint.
Related reading
- The full playbook for sales and marketing SaaS. Where this fits in the ordered programme.
- Training data versus live retrieval. Why broad questions and constrained questions behave so differently.
- How Perplexity chooses its sources. The engine where constrained questions move fastest.
- What AEO is. The general mechanism, if this is your first piece of ours.
Ask the four vendor questions about your own company this afternoon.
Pricing, integrations, implementation time, and why customers leave. Four questions, four engines, logged out. It takes fifteen minutes and most teams find at least two answers that are wrong in a way that is costing them deals they never hear about.
When you want the constrained questions tested too, at a sample that means something, send us the domain for a free visibility check. You keep the transcripts either way.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.