How MetriFi measured its own GEO experiment, and what the result cannot prove.

A company that sells generative engine optimization (GEO) reporting that GEO worked on its own website should make any reader skeptical. So this article starts with how MetriFi set up the experiment, what it wrote down before anything shipped, and the limits of the result. The numbers come after.

The method is the same one MetriFi uses on bank and credit union websites, and it is the part worth copying.

Step 1: Measure the starting point

MetriFi tracks a set of questions a bank or credit union marketer might ask an AI assistant when looking for help with AI visibility. Four of them asked, in different ways, what generative engine optimization is and which companies provide it to banks and credit unions.

Before changing anything, MetriFi read the AI responses to those four questions.

•  Visibility was 1.25 percent. MetriFi was almost never named.

•  The slot was open. In 27 of 27 sampled responses to the three main questions, the AI assistant named specific GEO providers for banks and credit unions. MetriFi was in none of them.

•  The case studies were working. On a separate question asking for documented content-experiment results, MetriFi’s published credit union case studies were already being cited.

So MetriFi had proof content that AI models were using, but nothing that told them what the category was or who MetriFi served. That was the gap.

Step 2: Find the mechanism before choosing the fix

Reading the responses showed why other providers were named. The models were repeating what those providers’ own pages said: a service page stating plainly that the company does GEO for banks and credit unions, or a guide titled with the category and the audience. None of the named providers were household names. They had published a page that said what they do.

MetriFi’s own GEO page did not do that. It had feature icons, two statistics and two calls to action, and no definition of generative engine optimization.

A second signal changed the plan. MetriFi’s AI bot traffic data showed which pages AI assistants were fetching when they browsed on a user’s behalf. Over 30 days, the GEO page was the fourth most fetched page on metrifi.com, with 28 fetches by ChatGPT’s browsing agent. Neither published case study appeared among the 74 pages fetched. The page was already where AI assistants landed, and it had nothing on it to quote.

An earlier round of this analysis had recommended writing a new guide first and treating the GEO page as unproven. The traffic data showed that was wrong, and the plan was corrected. MetriFi kept the earlier reasoning in the experiment record rather than overwriting it.

Step 3: Write the hypothesis down before shipping

Before either change went live, MetriFi recorded the hypothesis in the experiment. In substance, it read:

If MetriFi fixes the GEO page so it defines the category and names its audience in quotable text, publishes a cross-linked guide that does the same at length, and ships both together, then visibility on the four questions rises from 0 percent toward the roughly 15 percent held by the providers most often named, because the responses show MetriFi has the proof content and the inbound attention but no category-identity text for a model to cite.

It also recorded a trade-off. Shipping the page fix and the guide on the same day meant the experiment could never say which half moved the number. MetriFi accepted that on August 28, 2026, before either change shipped. MetriFi is not a bank or credit union, so isolating each change on its own site would not have told it much about client websites. Shipping both together tested the combined approach.

Writing both down in advance is what separates a measurement from a story told afterward.

Step 4: Ship a specific, limited change

Both halves went live on September 9, 2026.

The page. Six of seven planned edits went live that day; the seventh, adding the audience to the page title and meta description, followed later in the measurement window. The page gained a plain definition of generative engine optimization in body text, a headline naming banks and credit unions instead of the abbreviation “FI”, descriptive link text on the case study cards in place of “Read More”, and a link to the new guide. It grew from 519 to 941 words at launch, deliberately short of a long service page so it would not compete with the guide.

The guide. A 2,031-word article, what generative engine optimization means for banks and credit unions, opening with the same definition used on the page, so a model finding the definition finds the same sentence twice.

The links. The page links to the guide and the guide links back to the GEO page, so an AI assistant that reaches either one can reach the other.

Step 5: Measure the same way afterward

The measurement window ran 28 days, from September 9 to October 7, 2026, using the same questions and the same method as the baseline.

These are the final figures for that window:

•  Which GEO platforms serve banks and credit unions: 5.0% before, 73.5% after.

•  What is GEO for banks and credit unions: 0% before, 88.2% after.

•  Which vendors improve AI search presence for banks and credit unions: 0% before, 20.6% after.

•  What is the best GEO tool for banks and credit unions: 0% before, 5.9% after.

•  All four together: 1.25% before, 47.1% after.

The experiment set out to move MetriFi from absent toward about 15 percent. Two questions went well past that, one cleared it, and one moved only a little.

What this result cannot prove

Which half worked. The page and the guide shipped together on purpose. The result shows the pair worked. It cannot say whether the definition on the page, the guide, or the links between them did most of the work.

That it would be this easy for a bank. MetriFi was competing against a small set of pages that define GEO for this audience. A credit union trying to be named for auto loans in its city faces far more competition. This is a result about MetriFi’s category, not a forecast for yours.

That every AI assistant behaves this way. These responses came from one AI provider, OpenAI. MetriFi has not yet shown the same result across other assistants.

That nothing else changed. There is no control group. AI models are updated over time, and some of any change can reflect the model rather than the page.

That the first reading was clean. The first measurement on September 9 was taken earlier in the day than the page edits finished, so it sits closer to the baseline than to the result.

That it will hold. These figures cover one 28-day window that closed October 7, 2026. They are a reading of that window, not a permanent position.

The same method, on credit union websites

MetriFi runs this sequence on bank and credit union websites: measure the questions members ask, read why others are named, write the hypothesis down, ship one specific change, and measure again the same way.

In MetriFi experiments for three credit unions (Northwest Preferred Federal Credit Union, Lone Star Credit Union and HFS Federal Credit Union), visibility on targeted local questions went from 0 to 100 percent. In MetriFi’s published write-up of its experiment for Zing Credit Union, visibility rose from 66.67 percent to 85.71 percent after the credit union addressed how it compares with competitors. MetriFi’s write-up for Pearl Hawaii describes how the credit union went from invisible to cited in AI answers about Oʻahu CD rates through structure rather than more content.

MetriFi publishes experiment results regardless of outcome, including the ones that lose.

What to copy

1. Take a baseline before you touch anything. Without it, any later number is a guess.

2. Read the answers, not just the score. The reason a competitor is named is usually on that competitor’s own page.

3. Check what AI assistants already fetch. A page they visit that has nothing to quote is the cheapest fix available.

4. Write the hypothesis and the trade-offs down first. Including the ones that weaken the result.

5. Report the limits with the result. A number without its method is the thing AI assistants, and your board, have learned to discount.

To see how your institution appears in AI answers today, talk with us.