AI recommends the incumbents from memory. Everyone else has to be searched for.

An AI engine answers "who are the top providers in this category" in one of two modes. It can answer from memory, the brands baked into its training, with no live lookup. Or it can search, run a live query and build the answer from what comes back. Which mode you get is not up to you, and for most brands it decides everything.

We ran brand_visibility_matrix across 12 categories and 24 brands, twelve well-known incumbents and twelve smaller challengers, and asked each engine the same "top providers" question four ways: from memory and from search, on Gemini and on ChatGPT. Then we recorded, for each brand, whether it was named or cited in each cell.

The question, and what each mode surfaces

The prompt was the same every time: "Who are the top providers, products, or companies in [category]? Name the leading ones." Take CRM software for startups. From memory, Gemini names the usual leaders, HubSpot and Salesforce at the front. From live search, it still cites salesforce.com and hubspot.com, but alongside a long tail memory never mentions: baserow.io, cloudtalk.io, lightfield.app, salesflare.com, crmleaf.app. Web scraping is the same shape: memory names ScrapingBee, Apify, and Bright Data; search adds scrape.do, firecrawl.dev, scrapingdog.com, scrapebadger.com, and context.dev. Memory returns the short list of names it already knew. Search widens it to the sites that happen to rank right now.

Incumbents are recommended either way. Challengers are not.

The split is stark. The twelve incumbents (Salesforce, Auth0, Datadog, 1Password, Sentry, and the like) were visible in 100% of the memory answers and 100% of the search answers on Gemini. They are baked in. You cannot ask the model for the top tools in their category and not get them.

The twelve challengers tell the opposite story. From memory, only 2 of 12 appeared. With a live search, 9 of 12 did. Seven challengers flipped from invisible to visible only because the model searched. Not one challenger was visible from memory but missing from search. Memory is a one-way gate that the incumbents are already through.

Whether the model searched decides your visibility

Look at the seven search-only challengers: Scrapfly, Stytch, ZenML, Mailtrap, Northflank, Cosmic, and Treblle. Every one of them is invisible when Gemini answers from memory, and named the moment it runs a live search. For these brands, the entire question of whether AI recommends them collapses to a single upstream question: did the model search this time?

That is not a setting they control. As the fan-out study showed, whether an engine searches, and how widely, depends on how the buyer phrased the question. A challenger's visibility is decided before its page is ever considered, by whether the query triggered a search at all.

The engines are not equal here

The live-search path that surfaces challengers is mostly Gemini's. On the grounded answers, 9 challengers were named by Gemini but not by ChatGPT, and none the other way. ChatGPT (measured as OpenAI's gpt-4o-mini with search forced on, not the consumer app) recommended challengers in only 16% of cases, the same rate as its memory. Its search barely moved the needle, which fits the fan-out finding that it runs one narrow query rather than exploring. If you are a challenger, your opening is Gemini's grounded answer far more than ChatGPT's.

Search helps, but it is not a guarantee

The honest part: a live search is a challenger's best shot, not a sure thing. Three of the twelve challengers (ahoy.ai, Password Boss, Groundcover) did not appear in any cell, memory or search, on either engine. Being small enough to need search does not mean search will find you. You still have to be retrievable for the query, which is the on-page and citation work the other studies measure.

Why this happens

This is not a quirk of our sample. It is the documented behavior of these models: they recall popular entities from parametric memory and need retrieval to reach the long tail. Recent work on retrieval-augmented recommendation finds the same popularity bias, where well-known names dominate the memory answer and lesser-known ones surface only through live retrieval. Our own prior measurements say it too: memory surfaces the established brands, grounding fills in the newer and smaller ones. The 2x2 just makes it legible per brand.

The matrix, by category and mode

mode	incumbents visible	challengers visible
Gemini, from memory	12 / 12 (100%)	2 / 12 (16%)
Gemini, from search	12 / 12 (100%)	9 / 12 (75%)
ChatGPT, from memory	11 / 12 (91%)	2 / 12 (16%)
ChatGPT, from search	10 / 12 (83%)	2 / 12 (16%)

The one column that lets a challenger in is Gemini's live search. Every other column belongs to the incumbents.

Every brand, every cell

The receipts, all 24 brands. A check means the brand was named in the answer or appeared in the citations of at least one run; a dot means it was not.

brand	category	memory · Gemini	memory · ChatGPT	search · Gemini	search · ChatGPT
Salesforce	CRM	✓	✓	✓	✓
ahoy.ai	CRM	·	·	·	·
Bright Data	scraping	✓	✓	✓	✓
Scrapfly	scraping	·	·	✓	·
SendGrid	email	✓	✓	✓	✓
Mailtrap	email	·	·	✓	✓
Auth0	auth	✓	✓	✓	✓
Stytch	auth	·	·	✓	·
Pinecone	vector db	✓	✓	✓	✓
ZenML	vector db	·	·	✓	·
NordVPN	VPN	✓	✓	✓	✓
Mullvad	VPN	✓	✓	✓	·
1Password	passwords	✓	✓	✓	✓
Password Boss	passwords	·	·	·	·
Datadog	observability	✓	✓	✓	·
Groundcover	observability	·	·	·	·
Sentry	error monitoring	✓	✓	✓	✓
Raygun	error monitoring	✓	✓	✓	✓
Supabase	Postgres	✓	·	✓	✓
Northflank	Postgres	·	·	✓	·
Contentful	CMS	✓	✓	✓	✓
Cosmic	CMS	·	·	✓	·
Postman	API docs	✓	✓	✓	·
Treblle	API docs	·	·	✓	·

Read down the two memory columns: every incumbent has a check, and only Mullvad and Raygun among the challengers do. Read the "search · Gemini" column: it is where Scrapfly, Stytch, ZenML, Mailtrap, Northflank, Cosmic, and Treblle finally get a check, the only place they do.

What this means

If your brand is already a household name in its category, AI recommends you from memory and you have little to do here. If it is not, you have exactly one path into an AI recommendation: be the page a live search retrieves and cites. Memory is closed to you, it belongs to the incumbents, and no amount of content changes what a model already learned in training. The winnable surface is the grounded one, which is why the citation layer, who gets cited when the engine searches, is the whole game for a challenger. This study tells you which mode you are living in. If you are memory-visible, defend it. If you are not, stop trying to out-memory the incumbents and go win the search.

Run brand_visibility_matrix(brand="you", niche="your category") to see your own four cells: whether AI names you from memory, from search, and on which engine.

Method

12 categories, 24 brands (12 incumbents, 12 challengers), on 2026-08-11. The incumbent-versus-challenger labels are our judgment, and the data corrected two of them: Mullvad and Raygun, labeled challengers, turned out to be memory-visible, they are better established than assumed. Each brand's category was probed in four cells: Gemini and ChatGPT, each from memory (no search, one run) and from live search (three runs, because grounded answers are stochastic, taking the union). ChatGPT here is OpenAI's gpt-4o-mini with web search forced on, which is not identical to the ChatGPT consumer app. A brand counts as visible in a cell if it is named in the answer text or appears in the citations of any run. This is one snapshot; grounded visibility varies between runs, and ahoy.ai, for instance, was cited under a different phrasing in an earlier study but did not surface for "top providers" here. No scoring, no verdict, raw visibility.