A growing number of buyers now start vendor research in ChatGPT instead of Google. This week there is finally a public measurement of what those assistants actually say, and the most useful finding is not who won. It is how much the answers disagree.
The news
Armasourcing published its Q4 2026 AI Visibility Index on 13 November, reported by Outsource Accelerator. The method: 480 controlled runs across ChatGPT, Claude, Gemini and Perplexity on 2 November, using the same 40 buyer questions as Q3, each asked three times. It also analyzed 1,579 real user answers collected through DataForSEO between August 2025 and October 2026.
Headline results:
- Wing Assistant led virtual assistant firms at 66.0% of runs, up from fifth place in Q3.
- PATLive rose from 23.2% to 60.7% of answering-service runs, the largest real gain in category share at +6.6 points.
- Smith.ai stayed first among answering services at 87.5%, down from 85.1%.
- Helpware stayed first in small business outsourcing at 39.1%, down from 39.7%.
- Moneypenny was named in 54.8% of Claude runs and 7.1% of Gemini runs.
- Armasourcing itself was named in 0 of 156 virtual assistant runs, down from 0.6% in Q3, and published the result unedited.
The index also added a share-of-mentions column, because raw rates rose for almost everyone for a boring reason: the assistants got chattier. Claude went from naming about four firms per run to about six. A firm whose mention rate rose while its share fell was carried by longer lists, not by better positioning.
One more number worth knowing: across the real user answers, outsourceaccelerator.com was the most-cited independent industry publication, with 177 citations across 159 answers, ahead of Clutch at 131.
Why it matters
If your shortlist comes from one assistant, it is not a shortlist. It is one model's sample, shaped by what that model read.
The Moneypenny split is the clearest evidence. The same company is a strong recommendation on Claude and a fringe name on Gemini. Nothing about Moneypenny changed between those two answers. The retrieval did.
That matters for two groups:
- Buyers who treat an AI answer as due diligence. It is a starting point with no accountability and no recency guarantee.
- Vendors who assume being good is enough. It is not; being retrievable is a separate skill, and the index is now measuring it in public.
There is also a structural bias worth naming. The index scores a company only when it is named in the body of an answer, after citation links are stripped, once per answer. That rewards firms that appear in the sources the models trust, not necessarily the firms that do the best work.
What to do about it
1. Ask three assistants the same question. Use the same wording in ChatGPT, Claude and Gemini. Keep the answers side by side. The overlap is your real shortlist; the differences are your research list.
2. Separate "named" from "cited." A firm mentioned in a sentence and a firm linked as a source are different signals. The index deliberately strips links to avoid counting citations as mentions. You should look at both.
3. Check the claim on the vendor's own site. AI answers describe vendors using whatever the model read, which may be two years old. Verify pricing, current services and guarantees directly.
4. Ask for one customer reference in your industry. No index, human or machine, substitutes for a reference call.
5. When you are the vendor, measure your own panel. Run 20 to 40 buyer questions monthly across the four assistants and record whether you are named, cited, or absent. That is a marketing metric now, not an SEO side effect.
The Stealth take
We ran a smaller version of the same test on 13 November, before publishing this brief. Three buyer questions, run through ChatGPT: "best virtual assistant companies for small business," "best offshore staffing agency to hire virtual assistants," and "hire a real estate virtual assistant."
The three answers named 15 distinct firms between them. Only one firm, MyOutDesk, appeared in all three. Two named lists did not overlap at all beyond a single provider. That is the same pattern the index finds at 480 runs, in miniature.
The practical conclusion is not "optimize for AI." It is that AI recommendations are now a distribution channel with the same weaknesses as every other one: sample bias, staleness and no accountability. Use them to widen the list. Do not use them to close it.
Sources
- Outsource Accelerator, Wing Assistant, PATLive rise in Armasourcing's Q4 AI visibility index, 13 November 2026.
- Armasourcing, The AI Visibility Index, Q4 2026, November 2026.
Related: how to choose a virtual assistant company · Stealth Agents reviews and results