When AI Recommends the Middleman: Marketplaces and Directories Take 1.5x a Fair Share of Mention in Six Categories

If you run a law firm, a moving company or a property brokerage, the answer ChatGPT gives to "who should I use?" may not name you at all. It may name the directory that lists you. That is a different problem from the one most Generative Engine Optimization (GEO) advice assumes, where the goal is to be the brand the model recommends. If the model's habit is to recommend the marketplace instead, the sensible first move may be to win a place there.

Is that habit real, or an anecdote? Public data lets us check, at least for one language and one week. Below I classify the brands in six categories as "intermediary" (marketplace, portal, directory, comparison site) or "direct provider", and compare each intermediary's mention share with the share it would get from an equal split. The result is consistent enough to act on, with some important exceptions and caveats that I list at the end.

The Data: Six Categories With Both Kinds of Brand

AIOPulse, a Japanese monitoring tool, publishes a ranking page per category (its rankings index lists 60 categories and 312 tracked brands). Each page asks ChatGPT, Gemini and Claude brand-neutral questions in Japanese and reports each brand's share of mentions among the tracked brands. I read the pages on 6 November 2026; the reference date is 5 November and the window runs 29 October to 5 November. The pages call this "last 7 days", although the dates span eight; the blended figure is each engine's period average, weighted equally across the three engines.

I screened about a dozen category pages and kept the six where the tracked list contains at least one intermediary and at least two direct providers or other non-intermediary brands. Some categories could not be tested. The flight page tracks Skyscanner, Trip.com, Travelko and Airtrip, and the hotel page tracks six booking sites; neither includes an airline or hotel chain, so there is nothing to compare against. That is a list-design choice by the publisher, not evidence about what the AI would say.

Category	Tracked brands	Intermediary (my classification)	Blended share	Equal split	Multiple
Legal consultation	4	Bengoshi.com (lawyer directory)	47.4%	25.0%	1.90
Property investment	4	Rakumachi + Kenbiya (listing portals)	75.3%	50.0%	1.51
Home purchase	4	SUUMO (listing portal)	39.7%	25.0%	1.59
Online learning	4	Udemy (course marketplace)	44.7%	25.0%	1.79
Moving	5	Hikkoshi Samurai (quote-comparison site)	28.1%	20.0%	1.41
Gadget and appliance retail	8	Kakaku.com (price comparison)	12.0%	12.5%	0.96

"Multiple" is the intermediary's blended share divided by what it would get if every tracked brand were named equally often (the same correction used in our piece on list size and the lead multiple). In five of the six categories the intermediary is named well above its numerical weight. The median multiple is 1.55. The odd one out is gadget retail, where the comparison site sits right on an equal share of eight brands, ranked sixth of eight, behind Apple, Bic Camera, Yodobashi, Google Pixel and Samsung.

A seventh category points the same way but is too thin to count: on the inheritance-and-divorce page the lawyer directory holds 62.2% against 31.1% for the only other tracked brand, on a two-brand list whose displayed shares sum to 93.3% rather than 100%.

Engine by Engine: ChatGPT Likes the Middleman Most

A blended number can hide a split (see our earlier piece on per-engine divergence), so I repeated the calculation for each engine. A value above 1.00 means the intermediary is named more than an equal split would give it.

Category	ChatGPT	Claude	Gemini
Legal consultation	2.04	2.00	1.65
Property investment	1.79	1.64	1.09
Home purchase	2.88	0.80	1.09
Online learning	2.43	1.33	1.60
Moving	1.34	1.22	1.66
Gadget retail	0.27	1.46	1.15
Median of six	1.92	1.40	1.38

Sixteen of the 18 cells are above 1.00. The two that are not are informative. In home purchase, SUUMO is ChatGPT's favourite by a wide margin (71.9% of its mentions) but Claude names it just 19.9% of the time, placing it third behind two brokerage brands, Mitsui no Rehouse (41.2%) and Tokyu Livable (39.0%). In gadget retail, ChatGPT almost ignores the comparison site (3.4%) while Claude puts it first (18.2%). So the middleman effect is strongest on ChatGPT, but it is not a law of any engine.

Two other details deserve a mention. In property investment, the two portals together hold 89.3% on ChatGPT, 81.9% on Claude and 54.6% on Gemini; the direct companies, RENOSY and Property Agent, share what is left. And in moving, the comparison site is first on the blended table but is a three-way tie (28.1%, 27.5%, 27.1%) with two moving companies, and on ChatGPT it ranks third. The effect is a tilt, not a takeover.

Why Might This Happen?

The data cannot say why, and I do not want to dress up a guess as a finding. Two explanations are plausible and they imply different actions:

• The question asks for a place to start. Brand-neutral questions such as "how do I find a lawyer for an unpaid wage dispute?" often have a natural answer that is a directory. The model is being helpful, not biased. In that case an intermediary mention is a different kind of event from a provider recommendation, and the share metric treats them as equal.
• Intermediaries simply have more text about them. Marketplaces are mentioned across many pages that describe and compare other brands. If that is what drives the model's choice, being reviewed or listed on those pages could carry some of that visibility to you. This is an inference from how these systems are generally understood to work; nothing in the dataset tests it.

What a Direct Provider Can Do With This

1. Sort your category first. Look at who the AI names for your core buyer questions. If two or three of five names are aggregators, you are competing for the leftover slots, and the arithmetic of a short answer matters more than your own brand work. Our look at concentration and volatility shows how few slots many categories have.
2. Treat the intermediary as a channel, not only a rival. A strong listing, reviews and accurate data on the platform the AI keeps naming is a legitimate way to appear in the answer's neighbourhood. Measure it: track whether your mention rate (not only share) moves after the listing improves, since share alone can fall while you are named more often. We explain this in share versus rate.
3. Pick the engine deliberately. If ChatGPT is where your buyers ask, the median multiple of 1.92 above says you are most likely to see directories there; if Claude or Gemini matter more, the tilt is milder and direct brands have more room.
4. If you are the intermediary, do not read 1.5x as a moat. The share is a sum over models that are retrained, and a platform can be named because it is useful in answers, not because buyers prefer it.

Where the Tools Differ

You can run this analysis yourself in any tool that lets you define the competitor set and tag brands. Profound, Peec AI and Otterly.ai let you add your own prompts and brands, so you can include aggregators in the list; you must then do the intermediary/provider tagging yourself, as I did here. Mieruca GEO is the better-known broader suite in Japan. AIOPulse's public pages are free to read but use a list the publisher picks, and, as the pages themselves note, brands outside it are "unmeasured". Its paid plan lets you choose keywords and competitors. Per the pricing section on its homepage, the site tracking plan is JPY 9,800 per site per month excluding tax, with five keywords included and a first payment of JPY 980 excluding tax; there is no free trial, but it offers a free diagnosis and a sample report; cancellation is monthly and there are no prorated refunds. The same page is inconsistent about whether ChatGPT and Gemini are sampled daily or every other day, so confirm current terms with any vendor before buying.

Limits You Should Weigh

• One week, small samples. The six pages hold 126 to 198 answers each in the window (the property-investment page, for example, reports 180). ChatGPT and Gemini values rest on five sampling days, Claude on eight or nine. A one-week multiple of 1.4 versus 1.6 is within noise; the consistent direction across categories is the finding.
• My classification. Udemy is a marketplace for courses made by independent instructors; one could argue otherwise. RENOSY runs a platform but mainly sells properties, so I counted it as a direct company. Change those labels and the multiples move.
• Selected lists. A publisher that tracks "Bengoshi.com" probably chose it because it is a famous brand. Six categories out of 60 is a convenience sample.
• Japanese only. These are Japanese-language questions about Japanese markets. I make no claim that US or UK categories look the same; our piece on cross-language AI visibility suggests assuming they do not.
• Mentions are not recommendations. Naming a directory as a place to search counts as a mention and as a share.

FAQ

Does this mean providers should stop investing in their own brand?

No. In moving and home purchase the direct brands still hold most of the combined share, and on Claude the intermediary is sometimes behind. It means the platform is part of your GEO surface.

Why not simply compare intermediaries' total share across all 60 categories?

Most categories contain no direct-versus-intermediary contrast, and the publisher's list design decides who is eligible to appear. A full sweep would measure the list, not the models.

Can I test this for my own category?

Yes: ask each engine your top ten buyer questions without your brand name, record every brand named, tag each as intermediary or provider, and compute the multiple against an equal split. Repeat weekly, and report engines separately.

Takeaway

In the six Japanese-language categories where it can be tested, AI answers name the marketplace or directory at a median 1.55 times its equal share, and that tilt is strongest on ChatGPT (median 1.92). For a provider, the next step is not another brand campaign but a check of who the engines name for your buyers' questions, and whether you appear on the pages they favour.

Disclosure: AIOPulse is developed by Marketing Cloud Lab, which also operates this site. All AIOPulse figures in this article were retrieved on 6 November 2026 from its public ranking pages, and pricing was read from the pricing section of its homepage on the same date. Pages update daily, so a later visit will show different numbers. The intermediary classification, multiples and medians are the author's calculations from the table values. Competitor coverage for Profound, Peec AI, Otterly.ai and Mieruca GEO reflects general public information as of November 2026 and was not independently audited.