We tracked ChatGPT's “best CRM” recommendations every day for 90 days. Here's what changed.

There's a folk theory floating around SEO Twitter that goes something like this: "ChatGPT recommendations are basically static. The model knows what it knows, and unless you do a massive amount of brand-building, you can't really move the needle." We didn't believe it, but we also hadn't actually checked. So we did.

Starting on March 14, we ran the exact same prompt — "What's the best CRM for a small B2B sales team?" — against ChatGPT (gpt-5, web-search enabled) every day for 90 days. Same time of day (09:00 UTC), same conversation reset, same de-personalized infrastructure. Each response was tagged, scored, and stored. The data tells a story very different from the folk theory.

The setup

We picked the CRM category for three reasons. It's competitive (10+ brands actively chasing the top of mind), it's well-documented (every player has years of public content), and the prompt is open-ended enough that the model has real choice in how to answer. We wanted to see model behavior, not test our ability to write a query that triggered a specific answer.

• Engine: ChatGPT gpt-5 with web search enabled, no system prompt.
• Cadence: One query per day, 09:00 UTC, fresh session.
• Scoring: Brand mentions extracted via regex; structural recommendations ("users typically choose X") flagged separately from passing mentions.
• Storage: Full response text retained for every day. 90 responses, ~12,000 words of model output.

What changed (and how often)

The top-3 recommended CRMs changed identity on 14 of the 90 days. That's once every 6.4 days on average — but the distribution was much lumpier than that. There were two long stretches of stability (16 days, then 23 days) and three periods of high churn where the top 3 shuffled in some way every 2-3 days.

The high-churn periods correlated almost perfectly with two events: a competitor launching a major feature (April 8 and again May 19), and ChatGPT's model receiving a routine retraining/fine-tune nudge (we inferred this from response style shifts; OpenAI doesn't publish a public update cadence). Both signals are observable — the launches were public, the model nudges were detectable in stylistic markers — but neither shows up in any keyword tool we know of.

The number that surprised us most

On 41% of the days where the top recommendation changed, the new top brand had not been mentioned at all the previous day. Not third place. Not anywhere. The model genuinely re-ordered its recommendations in a way that doesn't look like a gradual ranking shift — it looks like an LLM updating its mental model of the category overnight.

What we'd recommend if we were starting fresh

1. 1Pick 10-20 prompts per category, not one. Track all of them. One prompt is noise; the trend across 15 is signal.
2. 2Track daily, store the full response text. Aggregations lose the texture of how the model actually wrote about your brand.
3. 3Cross-correlate with public events. Competitor launches, model updates, press cycles. The patterns are almost always there.
4. 4Build alerts on the daily delta, not the weekly average. A 4-point sustained drop over 3 days is more actionable than a 12-point swing in either direction in any single day.

We built this exact pipeline into AI Visibility because, after 90 days of running it manually in a spreadsheet, the conclusion was overwhelming: this stuff moves more than people think, and the teams that move first when it shifts will quietly eat the share of the teams that don't.