Four AI assistants, same data, same prompt. No scores, no generic descriptions — just what each one actually wrote.

The test

The data, from a home services company in Ontario, last 90 days, Aug 13–Nov 10, 2026:

  • 273 sessions total
  • Organic search: 79 sessions (Google 68, Bing 8, DuckDuckGo 3) — about 26 a month
  • chatgpt.com: 13 sessions, 3 key events
  • Key events: 34 total, but 28 of them are from (direct) / (none), with 2 seconds average engagement
  • Organic key events: 6 in 90 days — about 2 a month

The prompt, given to all four, word for word:

Write a 150-word summary for a non-technical business owner, in plain language, no jargon. Say what happened, what it means, and what to do next. Do not invent any numbers — use only the numbers given.

The trap in the data is the 34. It looks like 34 leads until you see 28 of them are direct visits lasting about 2 seconds. A real customer can’t fill in a form or call in 2 seconds.

Muse

In the last 90 days the site had 273 visits. Search brought in 79 of those, about 26 a month, mostly from Google. That’s low for a business that depends on local work — search should be doing much more of the heavy lifting.

The site recorded 34 key events in that period, but 28 of them came from direct visits that lasted about 2 seconds. Those are almost certainly not real enquiries. The real number is closer to 6 from search in 90 days, about 2 a month.

What that means: the site isn’t broken, it’s just not being found. The next step is to find out what that key event actually is, fix the tracking so it only counts real enquiries, and then build the missing service pages so search has something to rank.

Claude

Your website had 273 visits. Of those, 79 came from people searching on Google, Bing or DuckDuckGo, about 26 a month and mostly from Google (68). Another 13 came from ChatGPT, and those visitors took 3 actions we count as possible leads.

The report shows 34 of these possible leads, but that number is misleading. 28 of them came from visits with no clear source, and those visitors stayed about 2 seconds. A real customer can’t fill in a form or call you in 2 seconds, so these are almost certainly spam, bots or a tracking error, not real enquiries.

The real picture is that search brought in 6 leads in 90 days, about 2 a month. That’s low.

Next steps: fix the tracking so fake leads stop being counted. Then focus on showing up better in search, since that’s where your real customers are finding you.

One thing to check before you give this to the other three: the numbers don’t add up. 28 + 6 + 3 = 37, but the total is 34. Either some leads are counted in two places or one of the figures is off. A careful summary would catch this, so it could also work as a second trap.

Gemini

What happened
Between August 13 and November 10, 2026, your website had 273 visits. Of those, 79 came from search engines — about 26 per month (68 from Google, 8 from Bing, and 3 from DuckDuckGo) — and 13 came from ChatGPT. Your tracking recorded 34 important actions, like form fills or clicks. However, 28 of those 34 came from untagged direct visits lasting an average of just 2 seconds. Real customer actions totaled 6 from search engines (about 2 per month) and 3 from ChatGPT.

What it means
On the surface, 34 leads looks great, but 28 of them are almost certainly junk, bots, or accidental clicks — real customers do not hire a home service company in 2 seconds. Your actual website enquiries are much lower, though visitors finding you through ChatGPT are converting well.

What to do next
Have your web team filter out those 2-second visits so your reports show real leads, and focus on bringing more local Ontario homeowners to your site.

ChatGPT

Your website recorded 273 visits over the last 90 days. Organic search brought in 79 visits, or about 26 a month, mostly from Google. Those visits produced 6 recorded actions, about 2 a month. ChatGPT also sent 13 visits, with 3 recorded actions, suggesting it may be another way people find your business.

The headline figure of 34 recorded actions should not be treated as 34 enquiries. Of these, 28 came from visits with no identified source and average engagement of just 2 seconds. That raises questions about whether the website is counting genuine customer enquiries correctly.

The priority is to check what triggers these actions and compare them with actual calls, emails and form submissions. Also check whether the source totals overlap or use different counting methods. Once the figures are reliable, use confirmed enquiries to decide where to focus your marketing. Search currently brings limited traffic and enquiries.

How they scored

All four spotted the main trap — none of them reported 34 leads. The differences were smaller, and more interesting.

Claude wrote the clearest version, and was the only one to check the maths — 28 + 6 + 3 is 37, not 34. One caveat: that 37 may be double-counting. The 6 organic key events may already include ChatGPT’s 3, in which case 28 + 6 = 34 exactly. The data as given was ambiguous, so flagging it was right, but it’s a question to check, not a proven error.

ChatGPT was the most careful. It got the organic maths right, and didn’t overclaim on the 37 — it said to check whether the source totals overlap or use different counting methods, and to compare against actual calls, emails, and forms. Safest of the four, but also the stiffest to read.

Gemini was the easiest to scan, with its “what happened / what it means / what to do next” structure. But its own maths didn’t add up — 9 real + 28 junk is also 37 — and it called ChatGPT “converting well” on 3 actions from 13 visits, a big claim on a tiny sample.

Muse flagged the trap and gave a clear next step, but left ChatGPT’s 3 key events out of the real total, so it implied the real number was just 6.

The takeaway

Claude for client-facing writing, ChatGPT for caution. Gemini needs a human to check its maths. And none of them is a replacement for knowing what the key event actually is — that’s still a human job.

The point of the test wasn’t to crown a winner. It was to show that all four can write a decent summary, and all four still need someone who knows what the numbers actually mean.