Ask AI like a person and it names a quarter fewer vendors

What we did

The two sets carry the same 67 buying intents, in the same proportions, across the same three sectors. Only the phrasing differs. One example pair:

search-style: cheapest NDIS practice management software in Australia

conversational: “Got quoted $480 a month for NDIS software and nearly fell off my chair, there’s four of us working out of one small office in Geelong. What’s the cheapest thing that still does client records, notes and invoicing?”

Both sets were put to four AI systems — ChatGPT, Perplexity, Google AI Overviews and Google AI Mode — three times each, between 13 and 14 August 2026. 3,504 answers were analysed. Questions that name a vendor were excluded from both sides, because a vendor named in the question is named back almost every time and the two sets did not carry equal numbers of them.

What changed

Finding one

AI names a quarter fewer vendors

A 26% drop. Bootstrapped at 150 questions per side, so neither set has more question variety than the other; the confidence bands do not overlap.

Finding two

One answer in seven names nobody at all

Two and a half times as often. This is the finding with the sharpest edge: for one realistic buyer conversation in seven, no software vendor is recommended at all.

Finding three

The answer stops being a list and becomes advice

This is why. Of the conversational answers that named nobody:

• 36% set out buying criteria instead of names
• 11% said it depends on the organisation
• 8% asked the buyer a clarifying question

They are also shorter — a median of 2,683 characters against 3,971 for a search-style answer that named nobody. Not less to say. Different things to say.

One answer, verbatim, mid-sentence:

“The top three reputable NDIS software options in Australia span different organizational strengths and sizes:”

— and then, instead of the three names, a question about how large the provider is.

Ask like a search engine and AI gives you a shortlist. Ask like a person and it interviews you.

Finding four

It happens on Google too

We expected this to be a chatbot effect. It is not.

The shortening is universal — every surface we measured, including both Google products, names fewer vendors when the question is phrased conversationally.

The refusal to name anyone is not. That behaviour is concentrated in the conversational assistants: ChatGPT declines to name a vendor in nearly a quarter of conversational answers, Perplexity in nearly a fifth, while the Google surfaces stay under 10%. Google still gives you a list. It just gives you a shorter one.

Finding five

AI reads less, and more often reads nothing

A conversational question is twice as likely to be answered without the AI reading anything at all — from what the model already knows rather than from the live web.

Finding six

But the front of the list does not move

Among answers that did name someone, the vendors appearing in the first three positions are close to identical under both phrasings — the leaders hold their place to within a few percentage points.

And the total number of distinct vendors named across the whole study is unchanged: 75 under search phrasing, 76 under conversational. No vendor is locked out of the conversational answer. They are simply named less often.

So this is not a reshuffle and it is not concentration into the leaders. Measured across every vendor tracked, total visibility falls about 20%, while the share held by the top five is unchanged. Everyone gets less, because the list gets shorter and sometimes does not arrive.

Why it matters

Every vendor optimising for AI visibility today is optimising for the search-shaped answer, because that is the answer they have seen. It is the one that appears when you test your own category by typing your own keywords.

The conversational answer is a different contest. It has three places instead of eight, and one time in seven it has none — the AI asks a question back, or explains what to look for, and no brand is mentioned.

Two things follow, and neither is obvious from a keyword test:

1. Being on the list is not the same as being in the answer. A vendor that reliably appears seventh in a shortlist has a real search-style score and very little presence in a conversation.
2. The 15% is not a gap in the market — it is a gap in the answer. In one buyer conversation in seven the category is discussed and nobody is recommended. That share is not currently contested by anyone.

Method and limitations

• 400 questions, 4 AI systems, 3 runs each, 13–14 August 2026, Australian location signal. 3,504 answers after excluding vendor-naming questions.
• Vendor mentions are matched on brand names with word-boundary rules and case-sensitivity on names that collide with ordinary English. Brand names appearing only inside a web address are counted as a source, not as a naming.
• Confidence bands are 250–300 bootstrap resamples drawn at equal question counts on both sides. Equalising answer counts alone is not sufficient, and an earlier draft of this analysis reported two effects that did not survive the correction: a difference in the number of distinct vendors named, and a set of vendors that appeared to be absent from conversational answers entirely. Neither is real.
• The conversational set is written by us, not sampled from real buyer transcripts. It is calibrated against published measurements of real assistant prompts (median length and first-person rate) but it remains a constructed instrument.
• This measures what AI answers, not what buyers do next. We make no claim about how often Australian buyers phrase questions each way.
• Models. ChatGPT via the OpenAI API on gpt-5.1; Perplexity via its API on sonar; Google AI Overviews and Google AI Mode via DataForSEO, where there is no model to select — the answer is whatever Google served that day. The model is part of the instrument: changing one changes which vendors are named, so it is fixed for an edition and stated here.