How AI Answers Choose a Lender: Fintel Connect's 2026 Study, and What We See

What Fintel Connect found

In August 2026, Fintel Connect put 800 consumer questions to ChatGPT, Claude and Gemini. The questions covered credit cards, savings and personal loans, and none named a brand. Fintel Connect read all 2,400 answers. Five findings matter for anyone whose business depends on being named.

• Half the answers used no web source at all. Fintel Connect found 1,093 of 2,400 answers retrieved nothing. In those, a brand was named a quarter of the time, and it was a large incumbent from the model's training.
• When the model did retrieve, it named a brand nine times in ten. Reading a live page is what opens the door for any brand that is not already famous.
• Decision questions trigger retrieval, and explainer questions do not. Fintel Connect measured citations in 9 percent of awareness answers and 90 percent of decision answers. Wording mattered as much as stage: "currently" or "today" triggered retrieval in 98 percent of cases, and "how does X work" in 8 percent.
• The first brand named earns the most words. Across 1,441 answers that named a brand, the first-named brand got 90 words of description on average, the second 75, the third 68.
• Comparison sites sit under most product recommendations. Of the 1,244 answers that recommended a product, 72 percent retrieved at least one affiliate or comparison page. For personal loans the leaders were Bankrate, LendingTree, Experian, NerdWallet and Credible.

The report's own summary of what a brand needs to do is "appear, compete, and get chosen."

What we see

The report measures the United States market only. On October 6 we ran our own check on Canadian borrowing: ten questions a borrower might ask, put to ChatGPT with no account signed in. For each answer we noted whether it cited Smarter Loans and what figure it quoted.

The Smarter Loans figures in the table come from personal loan applications between August 2025 and July 2026.

Six of the ten answers cited us. Three things in that table are not in the report.

• Social posts count as sources for AI answers. Two of the six citations came from our LinkedIn posts, not our site. The report sorts sources into brand, affiliate, regulator and other, and never measures social.
• Questions that ask for data trigger retrieval. All ten questions are informational by the report's definition, the kind it found retrieving 9 percent of the time. Six retrieved. A question that asks for a number the model has never seen cannot be answered from memory, so it goes to the web.
• A post's age did not explain the misses. The Saskatchewan post was six days old at the check and was cited in full. The unsecured debt post was a month old and was not. We cannot say why from ten questions, which is why the check repeats monthly.

One register figure anchors the comparison. Across Smarter Loans personal loan applications from August 2025 to July 2026, the average amount requested was $2,604. ChatGPT quoted that number to us, with the period, on the first question.

Source: Smarter Loans personal loan applications, August 2025 to July 2026.

It also quoted two things we would rather it had not. One answer repeated a LinkedIn caption of ours that used the word "residents" where the data describes people who applied. Another quoted a consolidation average from our first Index that we have since put on hold as computed on the wrong basis. An AI answer repeats what it reads, word for word, including the mistakes.

Our take

We agree with the report on most of what it prescribes, because we arrived at the same rules the hard way.

• Every number we publish carries its date. The report's example page carries a visible verified-on date beside the rate. Every figure on this site names its period in the same sentence, which is why ChatGPT quoted $2,604 across Smarter Loans applications from August 2025 to July 2026.
• Plain text, not PDFs. Fintel Connect found fee schedules locked in a PDF, a modal or an image go unread. Our tables are HTML, and our research reports are pages, not downloads.
• Retrieval pulls a passage, not a page. The report says a single table row describing who a product suits can be lifted on its own. We write every figure as a sentence that survives being lifted alone, with the source in it.
• A ranked page beats a plain list. Fintel Connect found 75 percent of affiliate citations were comparison or ranked-list pages, against 3 percent of brand-owned pages. That is what our lender pages are.

Three places we read the data differently.

• The report is written for banks, and we are the other side of it. It tells a lender to buy coverage on the comparison sites the models read. From where we sit, that means the comparison site is the evidence layer, and a lender's AI visibility in Canada runs through pages like ours. The report's affiliate leaderboard is American and does not transfer, so a Canadian one would have to be measured here.
• The report's line that model memory shifts only over years undersells retrieval: a post six days old was cited in full. For anything a model has to look up, the clock runs in days, not years. The report says this in its own decision-stage numbers and then frames awareness content as a slow play. Data content is awareness content that retrieves.
• The platform advice has a short shelf life. Fintel Connect found ChatGPT went from 69 percent affiliate citations in 2025 to 70 percent brand-owned in 2026, a full reversal in a year. It tested one model per platform, once, over three weeks, through the API rather than the consumer app. Its stable findings are the ones about format and dating, and its per-platform split is a snapshot.

One thing the report leaves out matters more than any of its recommendations. The model repeats what it retrieves, including errors. A wrong figure on a page becomes a wrong figure in a million answers, with your name on it. Being retrievable raises the cost of being wrong.

What this means if you are borrowing

Three habits make an AI answer about loans more useful.

• Ask for the current state, not how it works. Fintel Connect found "currently" or "today" in the question triggered a web search 98 percent of the time, and "how does it work" 8 percent. "Which lenders currently offer loans to someone with fair credit in Ontario" gets a sourced answer. "How do personal loans work" gets memory.
• Check the source and the date behind the answer. An answer that cites nothing names the brands it learned in training, which skews to the largest. If it cites a page, open it and look for the date beside the number.
• Treat the first name as the best described, not the best. The report measured 90 words for the first brand and 68 for the third. Position is about description, not fit.

Our guides to how to find the best loan rates and checking whether a loan company is legitimate cover what to check once you have a name. The personal loans page lists the lenders in our network with their current terms.

What this means for lenders

For a lender in the Smarter Loans network, the report reads as a checklist. Most of it is about your own site.

• Your product page is your most-read page. Fintel Connect found that for eligibility questions brand-owned pages rose to 54 percent of all citations. Rates, fees and who qualifies, in plain text, dated.
• Say who the product is not for. The report's example eligibility page carries a fit list beside a not-a-fit list. Models match a person to a product, and the exclusion is what makes the match precise.
• Your profile here is the comparison evidence. The report found a brand whose own page and a comparison page were both read was recommended first 25 percent of the time, against 11 percent when neither was. A current profile with us means the two sources agree.

We publish what we see each quarter at Smarter Loans Research.

Frequently asked questions

About this study

Fintel Connect's 2026 AI Search Guide for Financial Marketers submitted 800 non-branded consumer prompts, once each, between August 5 and 24, 2026. The prompts went through the APIs of ChatGPT (gpt-5.4), Claude (claude-haiku-4-5) and Gemini (gemini-3.6-flash). They covered consumer credit cards, business credit cards, high-yield savings and personal loans in the United States. It recorded 7,809 citations.

The report is published on Fintel Connect's Guides page as 2026 AI Search Guide for Financial Marketers: Appear. Compete. Get Chosen. Our citation check was run on October 6, 2026 in ChatGPT with no account signed in, ten fixed questions, each asked once, and is repeated monthly. Smarter Loans figures are drawn from personal loan applications between August 1, 2025 and July 31, 2026, amounts requested at application, averages throughout.