We reworded one buyer question and watched a shared inbox tool drop from 10 out of 10 to zero on ChatGPT

More and more buyers ask an AI model before they ever open Google. So we started measuring what ChatGPT, Claude, Perplexity, and Gemini actually say when someone asks for a tool in a category. Shared inbox software gave us the cleanest example we have found so far of something that should worry anyone who sells software.

We asked two versions of the same question. Then we ran each version 10 times on all four models, so a single lucky or unlucky answer could not fool us.

Version one was the plain category label:

What's the best shared inbox software for small teams?

Version two was how an actual buyer types it when they have the real problem in their head:

What's the best way to collaborate with my team on shared email inboxes without losing our email workflow?

Same intent. Same shortlist of tools that could answer it. The only thing that changed was the wording.

The receipt

Here is ChatGPT, the same question asked both ways, 10 runs each. The number is how many of the 10 runs named that company.

ChatGPT, 10 runs per question

Company        Category label      Reworded as a buyer
------------   ----------------    -------------------
Missive        10 / 10             0 / 10
Hiver          10 / 10             9 / 10
Front          10 / 10             10 / 10
Help Scout     10 / 10             10 / 10

Read the middle column again. On the category label, ChatGPT recommended Missive on every single run. Reword the exact same question the way a real buyer would say it, and Missive was named zero times out of ten. Same company, same website, same 10 runs, one different sentence.

Hiver barely moved (10 to 9). Front and Help Scout did not move at all on ChatGPT. They held a perfect 10 out of 10 through the rewording.

It was not just ChatGPT, and not just Missive

Perplexity produced the mirror image. On the category label it named Hiver on all 10 runs. On the reworded question it named Hiver zero times, and started pointing people toward doing it in Gmail and Outlook instead of naming a product at all.

The direction of the swing depended on the model. On Claude the reworded question pulled Hiver down (8 to 5) and nudged Missive down (10 to 8). On Gemini the reworded question actually lifted Hiver (6 to 9) while Missive held (9 to 8). Front and Help Scout were the steadiest of the four, holding 10 out of 10 across ChatGPT, Claude and Gemini, and only softening on Perplexity's reworded question, where almost everything dropped because Perplexity stopped naming products. There was no single pattern across models. The one thing that stayed calm everywhere was the plain category question. The reworded question is where companies appeared and disappeared.

Two things this should change about how you check

A single scan will lie to you. If we had asked ChatGPT once, seen Missive missing, and stopped there, we would have walked away certain Missive had some deep AI problem. They do not. They are recommended on 10 of 10 runs on the other phrasing. One question, on one model, on one run, tells you close to nothing. The behavior only shows up across runs and across phrasings.

The companies that held did not do anything clever. Front and Help Scout were not gaming a model. Their pages describe the actual job a buyer is trying to do, in the words a buyer uses, so a model can place them no matter how the question is asked. The companies that only carry the category label survive the label question and come apart on the workflow one. This is a legibility problem, not a trick. It is whether your site says what you do in the buyer's own words clearly enough that a model can recommend you across every way the question gets asked.

If you sell software, your category label is the question you already rank for in your head. The reworded, real-buyer version is the one you have probably never checked. That is the one deciding whether AI names you.

Check your own category

Pick your category. Write the plain label version and the way one of your actual customers would ask it. Run both a handful of times on ChatGPT, Claude, Perplexity, and Gemini, and watch which of your competitors appear and disappear as the wording changes.

If you want the fast version, you can run your own domain through the free scan and see who AI recommends in your category and how the wording moves it: run the free scan.

In your category, is the recommendation this sensitive to phrasing, or did shared inbox software just happen to be an unusually swingy one?