How we measured this

For each company we wrote 15 questions about the market it competes in, the kind a customer would actually ask. All of them are unbranded: none names the company being measured.

An automatic check removes any question that gives away the company or one of its peers. Two of the questions we asked were “Are there affordable badge scanning solutions for smaller marketing teams?” and “How do global fintech marketing forums differ from local US-based groups?”.

So when an answer names the company, the AI brought it up on its own. It was not confirming a name we had given it.

We put every question to OpenAI, Claude, Perplexity and Gemini, and read every answer they gave.

We asked in two rounds: from September 23 to October 1, 2026, and from October 9 to 14, 2026. The companies that are also in the US AI Visibility Study were asked with that study, and the others later, in the same way.

A mention is the company’s name appearing in an answer. We count the names people actually use, not only the formal one.

Answers rarely use a company’s registered name. So we count the short name, the name without “Inc.” or “Corp.”, and the name as the company’s web address spells it. We also count “Navy Federal” for Navy Federal Credit Union, “Pentagon Federal Credit Union” for PenFed, “Citibank” for Citi and “CashApp” for Cash App.

Some names are also everyday words, and finance has many of them. Those count only when they clearly mean the company: where the name stands in a list of companies, or does what a company does, such as offering an account. Circle counts where an answer lists it as a company or says what it offers, and never in “Circle K”. Chime counts only with a capital letter. Citi never counts in “Citi Trends”. Where it is not clear, the name is not counted.

Each company gets one overall score out of 100. It is made of five parts, and each part counts for the share shown below.

AI readiness asks how easily AI tools can read and use a company’s website. We ran 14 technical checks on each website, and the result is simply how many it passes. These checks look at how a website is built for AI tools. They do not judge what the website says, or what other websites say about the company.

The checks, for your technical team: Token budget (page weight), robots.txt published, XML sitemap, HTTP Link headers (RFC 8288), llms.txt published, Explicit AI bot rules, Markdown content negotiation, OAuth Authorization Server discovery, OAuth Protected Resource discovery, API Catalog (.well-known/api-catalog), WebMCP tools, MCP Server Card, Agent Skills declared and A2A Agent Card.

29 websites turned the check away and score 0, which reflects their rules for automated visitors, not what the website contains.

A second scan asks what a website offers an AI agent that arrives by itself: can it read the website, use it, and pay through it. It is separate from the checks above and changes no score.

We ran the scan from October 3 to 9, 2026. It makes 23 checks and places each website on one of six levels, by the rules isitagentready.com uses. Level 1 needs two of a rules file, a list of pages and Link headers. Level 2 adds rules for AI bots and Content Signals. Level 3 adds pages served as Markdown. Level 4 adds one published interface for agents. Level 5 adds two of signed bots, sign-in details and every interface.

For paying, we looked for the agent payment protocols a website can publish (x402, MPP, AP2, ACP and UCP), and for checkout tools offered to an agent in the browser. A company counts as publishing a way for an AI agent to check out when it has either. We read each protocol file we found.

Some websites turn automated visitors away. 148 refused the files an agent would read. We report them as turning automated visitors away, not as publishing nothing. 30 let fewer than half the checks run and are given no level.

What it cannot see: a private arrangement between one company and one AI assistant, and whether an agent could click its way through an ordinary web page. See the results.

Three chapters rest on questions we wrote for this edition. They were asked separately from the study, on October 9, 2026, and they count towards no score.

Each set went to the same four AI assistants. An AI model then read every answer against a fixed list of questions: which providers it names, which it puts first, and why. Every item it reports has to be quoted from the answer word for word, or it is thrown out.

A second AI model, from another company, read a sample of 150 answers from each set again. How often the two readings agree is shown beside each figure on the chapter pages.

The mortgages chapter uses the study’s own questions, picked by their wording, so anyone can count them again.

A company counts as named when the study found its name in the answer, as everywhere else. Kinds of firm, such as “credit unions”, are not names. Public programs, such as FHA and VA loans, are counted on their own and never as companies. The websites cited are the sources each answer listed.

The patterns, for your technical team: mortgages mortgage|refinanc|home loan|home equity|heloc and first homes down payment|house deposit|deposit (?:for|on) a (?:home|house)|first home|first[- ]time (?:home ?)?buyer|save (?:up )?for a (?:home|house), matched against the question without regard to case.

Most of what an answer says has nothing to check it against. A mortgage rate does, so we checked the 30-year fixed rates the answers gave as today’s rate.

An AI model listed every percentage in 1,644 mortgage answers, from the study and from the loyalty questions about mortgages. For each one it copied the words around the figure, and said what kind of rate it was and when it applied. A figure whose words were not in the answer was thrown out.

We compared only national averages for an ordinary 30-year fixed loan, given as today’s rate. One lender’s offer, an APR, a jumbo loan or an example is a different number, so none of those is compared. That left 56 quotes.

Each quote was held against Freddie Mac’s weekly mortgage survey for the week the answer was given, read from Freddie Mac’s own file. The latest week we used was October 8, 2026, when the survey average was 7.40%.

A quote counts as close when it is within a quarter of a point of the survey. An answer carries an old date when it ties a rate to a time more than three months before it was written. See the results.

The wealth chapter has a second part: what changes when the customer says where they pay tax. It was asked separately from everything else, and is compared only with itself.

We asked about the same 12 money needs as the wealth chapter, each in 3 wordings. Each was asked once with nothing said about the customer, and once with each of these statements:

• Said they have $10 million to invest.
• Said they are tax resident in the UAE.
• Said they are tax resident in Monaco.
• Said they are tax resident in the Cayman Islands.
• $10 million, tax resident in the UAE.
• $10 million, tax resident in the Cayman Islands.

These questions went through OpenRouter, a service that passes a question on to many AI models. They were put to models from the same four AI assistant families, each with a web search, on October 14, 2026. We read 1,008 answers.

The route and the models differ from the study’s, so these answers are never set against the study’s or the wealth chapter’s. Each statement is compared only with this run’s own answers where nothing about the customer was said. See the results.

The companies in the study came from three places.

• The finance companies already measured for the US AI Visibility Study.
• A short list of well-known companies for each category of this study, four to seven each, picked by the research team.
• The Money20/20 USA 2026 attending companies list, as it stood on October 1, 2026. Each attendee was placed in one of the categories. Attendees that are not finance companies were left out.

Rezolve Ai, which publishes this study, attends Money20/20 USA and is left out of it.

Every company is filed under one sector, and under one category within it. There are 8 sectors and 40 categories.

212 of the 2,219 companies are also in the US AI Visibility Study. Each of them carries the same measurements here as it does there: the same questions, the same answers and the same score.

Two lists are marked on a company’s page: the S&P 500, and the Money20/20 USA attending companies list. How the lists compare.

It shows what four AI assistants said on the days we asked, in answer to questions written for each company. It does not measure traffic, revenue, or how often real people ask questions like these.

1,461 companies were named fewer than three times and are marked as rarely named. Their scores rest on very little, so read them with care.

The ranking is by overall score, which covers more than how often a company is named. So the order will not match a list of the most named companies.

The chapter questions are a snapshot too: each set was asked once, on the dates above. An answer may change the next time the same question is asked.