Which automation tools do AI assistants actually recommend? Sweep of 28 September 2026

Every week we ask the big AI assistants the questions a real buyer would ask about automation, and we write down every answer: who got named, who got recommended, and which pages the assistant leaned on. This is the report for the sweep run on 28 September 2026. It covers 181 answers to 47 buyer questions across Claude.

We are Autoploy, and we run this measurement on ourselves as well as on everyone else. Our own score is in here, unrounded, including the weeks it is zero. That is the whole point: a number you can only trust if it is allowed to be bad.

The headline number

Change against the previous sweep: +0 vs 2026-09-21.

An unprompted question is one where the buyer never types a brand name. Those are the only ones in the headline. Questions where the buyer names a vendor in the question itself are reported separately further down, because folding them in would flatter everybody who is already known.

Who the assistants recommend when nobody names a brand

These counts are answers to unprompted questions in which each tool was named at least once. An answer usually names several tools, so the shares add up to well over 100 percent. Denominator: the 43 unprompted answers that did not name Autoploy.

Zapier is the tool taking the most demand we cannot yet reach: it appears in 28 of those 43 answers (65.1 percent). If you are building in this category, that is the incumbent answer you are being compared against before a buyer has heard of you.

The ghost rate: demand that gets described but not matched

Ghost rate is the number we watch hardest, because it is the honest size of the gap. A high ghost rate does not mean the assistants dislike you. It means they have never been given a reason to know you exist.

By kind of question

What each of those rows means, in plain words:

• Leaving a tool they already use: someone saying their current automation tool is too expensive or too fiddly and asking what to move to.
• Describing the job they need done: someone describing a job in their own words, with no tool named at all.
• Asking what is out there: someone asking what tools exist in this category.
• Asking for an AI that builds the automation: someone asking for an AI that can build and run the automation for them.

By assistant

Where the assistants got their answers

Across all 181 answers the assistants cited 1016 pages, from 307 different sites. 35 of the answers showed their sources at all. These are the sites they leaned on most.

Autoploy pages were cited in 0 percent of answers this sweep. If you want to know why a vendor is or is not in an AI answer, this table is a better clue than any ranking report: the assistant can only recommend what the sources it reads have written about.

What happens when a buyer types the name

This is the entity problem, and it is worth publishing because it is not only ours. Other companies share the Autoploy name, so an assistant asked about us sometimes answers about somebody else entirely: 1 of the 4 answers above were about a different company, and 0 unprompted answers hit the same confusion. Any young brand with a name somebody else also uses has this problem, and mostly does not know it.

How this was measured

• A fixed set of buyer questions is asked to each assistant with web search turned on, once per sweep, and every answer is stored whole.
• 181 answers were recorded this sweep across 47 questions and Claude. 134 probes failed to return an answer and are excluded from every number here.
• Only questions where the buyer names no brand count towards the headline score. There were 43 such answers.
• A mention only counts for a vendor when the answer or its citations point at that vendor. For us that means an autoploy.ai signal, because unrelated companies use the name elsewhere. Without that rule this report would have flattered us and been worthless.
• The exact wording of the questions is not published, and part of the question set is held back even from our own team, so no one can quietly aim at the test instead of the market.
• Competitor counts are one per answer: naming a tool three times in one answer counts once.

Bias statement, in plain words: we are a vendor in this category, we chose the questions, and we run the measurement. What we can promise is that the method is the same for everyone in the table, that our own number is reported exactly as measured, and that we publish it in the weeks it is embarrassing.

Who publishes this

Autoploy is an AI automation platform where you build automations by talking to Claude, and they run hosted and heal themselves when they break.

This report is produced by an automation running on Autoploy: it reads the week of probe results, counts them, writes this page, publishes it, and then checks the page is really live before recording that it published anything. Nobody types these numbers.

Data: our own probe sweep of 28 September 2026. Published 5 October 2026. Previous report: the sweep of 21 September 2026. A new report follows every sweep. Figures may be quoted with attribution to Autoploy.