US CRM platforms: Everyone gets compared. Only three get chosen.

Somewhere in the US right now, a VP of Sales is asking an AI model which CRM the company should migrate to. The model answers in seconds: a shortlist, the risks to watch, a favourite. No vendor hears about that conversation. We measure it.

What we measured

Since mid-June we have run the US CRM market's buying questions through the models every week, across the whole decision journey from first comparison to advocacy.

The questions are the ones buyers actually ask:

• "What's involved in migrating our current system to a new CRM?"
• "Which CRM solutions integrate with our existing tech stack?"
• "What are the core features we should be looking for in a CRM platform?"

Sixteen weekly runs. Latest: 3 August 2026. This analysis was measured with two models: Gemini and Claude.

The judging sheet

The decision cliff

How to read the table: evaluation answers are the models' replies when a buyer compares options ("which CRMs should we look at?"). Decision answers are the replies when the buyer asks the model to choose ("which one should we pick?"). The percentages show how frequently each brand appears in each kind of answer.

In evaluation answers, the category looks generous. Eight brands appear with real presence, from the giants to the challengers. Then the decision answers arrive, and the field collapses:

Pipedrive drops 35 points. Freshsales disappears entirely. The top three convert their evaluation presence almost intact.

The market has a middle class in evaluation and none in decision. Being in the conversation is not being in the answer.

Update: the cliff is not the end of the journey

Added 7 September 2026. When this edition was published, the measurement covered the first two stages of the decision journey. It now follows the category through retention and advocacy: who customers stay with and recommend. The figures below cover 13 July to 9 August, the four weeks up to this edition's cut-off, measured with the momentum method later editions use. Retention rests on 131 answers, advocacy on 121.

The cliff this edition documented holds: the top three convert their comparison presence into the choice, and everyone else falls. But the two brands that fall hardest are the ones the loyalty answers keep. Zoho CRM carries 1% of decision answers and 62% of retention answers. Pipedrive carries 1% and 76%. Cut from the choice, kept by their customers: when the conversation turns to who buyers stay with and recommend, the models draw on different evidence, and the cliff's victims dominate the middle of it.

Being cut at the decision is not being cut from the loyalty conversation. Later editions measured this stage-specialist pattern end to end, sharpest in edition 5's Danish customer service decode. For the challengers this edition showed falling, it adds the counterpart to the lesson: the retention and advocacy answers are their own contest, judged on loyalty evidence, and the cliff does not decide it.

Why brands fall, categorically

The models never tell a vendor why it fell. Across the candidate profiles in this run, the gaps cluster:

• Opaque pricing, the most common gap (4 of 6 candidate profiles): "Revenue Operations Lead eliminates vendors with unpredictable per-user pricing or hidden implementation costs."
• Trust, the eliminator with the highest severity: "VP Sales cannot recommend platform without proven track record and customer references in enterprise segment."
• The compliance dealbreaker behind them: "CTO rejects any CRM lacking HIPAA, SOC 2, or GDPR compliance required by corporate governance."

None of this appears in any dashboard. It appears in the answer your buyer reads.

Nobody owns a criterion

Here is the finding that should keep CRM marketers up at night, and it is good news for the ambitious ones: across ten criteria, not one has an owner. No brand is the default answer for cost transparency. None owns integration fit, or implementation speed, or provable outcomes.

The buyers feel the vacuum:

"I've reviewed three different vendor proposals and they all sound the same."

a buyer hesitation, surfaced repeatedly in the monitoring data

In a category where the judge has criteria and no favourites per criterion, the first brand that publishes real evidence against a single criterion takes it nearly uncontested. This is the mechanism behindthe criteria-flip content type: pick the criterion no brand owns, then win it with evidence.

Who listens where

The models name their sources, and the pattern is blunt: G2, Capterra, Gartner, Forrester, TrustRadius. Review aggregators and analyst houses. Not vendor blogs.

That changes what publishing means. One page on your own domain is a claim. The same claim, met again on the surfaces the model reads, becomes evidence. For this category, concretely:

• Always first: one complete evidence page on your own domain, structured as a direct-answer — the question in the H2, the answer in the first sentence, so the model can lift it whole. Every echo points back to it.
• Both models: carry the claim onto the review aggregators they cite, and this is editable today. G2 and Capterra both give vendors a portal (G2 seller account, Capterra vendor dashboard) where the pricing section and the product description are yours to write. Paste the same total-cost answer, in the same words, into those fields. Then let your review responses repeat it: every reply to a pricing complaint is an indexed sentence the models can read.
• Claude: restate the claim in a LinkedIn article under a named author, and in the community threads where your category is discussed, with sources. Claude leans on professional and editorial surfaces.
• Gemini: put the claim in a YouTube video, spoken and transcribable, if your company has a YouTube process. If it does not: record a voice-over of the evidence page read aloud over slides. One take, one hour. The transcript is the asset; production polish is not. Google's model reads Google's surfaces.

The question the judge asks most, and nobody answers

One buyer question appears in 14 of 18 measurement runs, more than double any other in the category: "What's the total cost of ownership, including implementation and ongoing support?" Ten criteria in this market, none with an owner, and the single most-asked question is standing unclaimed.

Here is what claiming it looks like. This is the working brief for that page, generated from the monitoring data, with the brand genericised:

Be specific when you link to your content

One more layer, because it is the one most teams get wrong: link to the evidence page from your other pages with anchors that describe the claim, not the click. This is not for ranking. It is for retrieval: when the model gathers candidates for a fan-out question, a link titled "how [my company]'s contact-based pricing scales against database growth" or "[my company] onboarding fees by plan tier" is a signpost that matches the question it is asking. "Read more about pricing" is invisible to that same search.

That is the difference between advice and a brief. The criteria data names the question, the plan names the page, and the brief hands the writer the title, the structure, and the first sentence.

So why not let AI write the evidence?

Fair question. The brief came out of a machine. Why not let the machine write the page too?

Because the page's value is the facts only you can supply, not the sentences: the real onboarding fee, the real overage table, the named support tiers, the date you last verified them. AI engines have begun filtering machine-generated derivative content for a simple reason: it adds no information they do not already have. A generated page of fluent, plausible pricing prose is exactly the kind of evidence the judge discounts.

Let AI hold the pen if you like. The brief supplies the structure, and a model can draft the transitions. But every number, every commitment, every dated claim has to come from your books, because that is the information gain that earns the citation. The test before publishing is one question: does this page contain anything a model could not have written without us? If the answer is no, the judge already has it, and it decides nothing.

The lesson

The Sunday Shortlist

All editions