Part 2: What 84 AI Visibility Experiments Taught Us About Credit Union Websites
Guest Editorial by Derik Krauss, Co-Founder of MetriFi
A year ago in Finopotamus, I shared how three credit unions went from 0 to 100% visibility in AI search. Northwest Preferred Federal Credit Union, Lone Star Credit Union and HFS Federal Credit Union each published content written to answer one specific question a member might ask an AI assistant, and each went from 0% visibility on its target question to 100%.
The most common question I got afterward was a fair one: was that luck?
Three wins make a good story. They do not make a pattern. So over the past year, my team at MetriFi kept running the same kind of experiment, on more credit union and bank websites, measured the same way every time. MetriFi has now run more than 90 generative engine optimization (GEO) experiments, nearly all of them on credit union and bank websites.
This article looks at the 84 that had finished as of October 22, 2026. This is what they show, including where the early lessons held up and where they need an asterisk.
The Short Answer
Across MetriFi's experiments, publishing new content and improving existing pages both raised visibility about 7 times in 10, by similar margins. What the page was mattered far more: long, explanatory resource pages did best, and edits to product pages barely moved.
What was measured | Average visibility change | Win rate | Experiments (n) |
New content published | +15.7 points | 71% | 65 |
Existing page improved | +17.8 points | 73% | 15 |
Long content (3,000+ words) | +26.4 points | 76% | 29 |
Medium content (500 to 3,000 words) | +11.0 points | 70% | 40 |
Resource page | +20.9 points | 75% | 36 |
Blog post | +14.4 points | 76% | 29 |
Product or service page | +3.8 points | 33% | 6 |
Best combination: new, long, explanatory resource page | +43.7 points | 88% | 8 |
Source: MetriFi GEO platform, finished experiments as of October 22, 2026. A "point" is one percentage point of visibility, the share of AI responses to a tracked question that name the institution. Win rate is the share of experiments where visibility rose at all.
Taken together, about 7 in 10 of the 80 MetriFi experiments that recorded an action ended with higher visibility than they started with.
How We Measured
The method is the one I described in Part 1, applied consistently.
A tracked question. Each experiment starts with a small set of questions a member might actually ask an AI assistant, such as where to refinance an auto loan in their city or which local credit union has the best certificate rates.
Visibility. MetriFi sends those questions to AI models and records whether each response names the credit union. Visibility is the share of responses that do.
Before and after. We read visibility before the content goes live, then again after it has been live long enough to gather a meaningful set of responses. The change between the two is the result.
What actually shipped. For every experiment, MetriFi records whether we published something new or improved an existing page, and for the page that went live, how long it is, where it sits on the site and what it is for.
Points, not percentages. Every figure here is a change in percentage points. Relative gains can sound large and say little: a credit union named in 1 of 100 AI answers that moves to 5 of 100 has improved 400 percent and is still missing from 95 answers out of 100. In points, that is a gain of 4.
Part 1 Lesson: Answer the Whole Question. It Held Up
In Part 1, the common thread across all three wins was long-form content that directly answered one specific question. MetriFi's larger dataset backs that up more strongly than I expected.
Long pages, 3,000 words or more, averaged a 26.4 point gain with a 76 percent win rate across 29 experiments. Medium pages, 500 to 3,000 words, averaged 11.0 points with a 70 percent win rate across 40.
The asterisk: this is not evidence that longer is always better. Cross-industry research on AI citations has found little relationship between raw word count and being cited. What the long pages in MetriFi's experiments had in common was that they fully answered one question for one market, and that takes words. Padding a page to hit a number is not the same thing.
Part 1 Lesson: Give AI A Page Built To Answer. It Held Up Too
Part 1 suggested that a credit union website needs two kinds of pages: pages for people, easy to scan and visually appealing, and pages for AI, text-based, objective and thorough. The data is clear on this one.
In MetriFi's experiments, resource pages, evergreen educational pages that are not dated posts, averaged 20.9 points across 36 experiments. Blog posts won about as often (76 percent, against 75 for resource pages) but by less, averaging 14.4 points across 29.
Product and service page edits averaged 3.8 points, and only 2 of 6 won. Six experiments is too few to be conclusive, but it matches what we see when we read the AI answers themselves: an assistant answering a "where should I" question cites pages that explain and compare, not pages built to sell one product. A product page can be excellent at converting a visitor and still give an AI model very little to quote.
New Lesson: New Or Improved Matters Less Than What the Page Is
A common piece of advice, and one AI assistants themselves repeat, is to update your existing pages first. In MetriFi's experiments, new and improved pages performed about the same.
Across 15 experiments that improved an existing page, visibility rose an average of 17.8 points and rose at all 73 percent of the time. Across 65 that published new content instead, the average was 15.7 points with a 71 percent win rate.
That gap is small, and 15 experiments is a small group, so I would not pick a side on it. The bigger differences in MetriFi's data come from what the page is: long pages beat medium ones by about 15 points, and resource pages beat product pages by about 17. If a page on your site nearly answers a member's question, improving it is a reasonable first move. If nothing on your site answers it, publish the page.
New Lesson: Purpose Matters as Much as Format
MetriFi also classifies each page by its primary purpose.
Primary purpose | Average visibility change | Win rate | Experiments (n) |
Guide (step by step) | +20.0 points | 90% | 10 |
Explain | +19.7 points | 67% | 30 |
Persuade | +16.7 points | 40% | 5 |
Compare | +13.0 points | 78% | 23 |
Step-by-step guides won most often. Comparison pages won often but by less. Persuasive pages had the lowest win rate: when they worked they worked well, but most did not.
The strongest combination in the dataset was a new, long, explanatory resource page: 43.7 points on average, with 7 of 8 experiments winning. It is also one of the smallest groups, so I read it as a strong pattern rather than a rule.
Not every new page wins, either. New, medium-length comparison pieces published as blog posts averaged a 19.7 point decline across 5 experiments, the only combination with a negative average. Five experiments cannot explain why, but they are a useful reminder that publishing something is not the same as publishing the right thing.
What This Data Cannot Tell Us
I would rather you trust these numbers for the right reasons, so here are the limits.
Individual experiments are small. A single experiment reads a limited number of AI responses, enough to see a large swing but not always a modest one. Pooling 84 experiments is how a pattern becomes visible. MetriFi keeps refining how it measures, and now sizes new experiments by the number of responses they need rather than by a fixed number of days.
A "win" means visibility went up. Any increase counts as a win, however small. The average change in points is the better guide to how much something moved.
This is not an A/B test. A website A/B test splits visitors between two versions of a page at the same time, which is why it remains the standard for proving cause. AI answers cannot be split that way, so each experiment compares a credit union's own visibility before and after a change. AI models are updated over time, and some of any change reflects the model, not the page.
Groups are uneven. Some rows rest on 29 or 36 experiments, others on 5 or 6. The small rows are directional at best, which is why the number of experiments is printed next to every figure.
None of this makes the pooled results meaningless. They are best read as consistent patterns across many institutions, not a guaranteed return on any single page.
What I Would Tell A Credit Union Marketing Team Now
Start from the question, not the page. List the questions your members actually ask an AI assistant about your products in your market, and check whether your site has a page that answers each one directly.
Fill the gap, new or improved. If a page nearly answers the question, improve it. If nothing does, publish one. In MetriFi's data, both raised visibility about 7 times in 10.
Make it complete and explanatory. Answer the whole question for your market, in plain language, with specifics a model can quote.
Let product pages do their own job. Keep them focused on converting visitors, and give AI assistants an explanatory page to cite.
Measure before you publish. A baseline taken before the page goes live is the only way to know whether it worked.
The Takeaway
Part 1 showed that a credit union can go from invisible to named in AI answers. Part 2 shows that it was not luck. Across 84 MetriFi experiments, the same few moves kept working: answer a real member question, answer it completely, and put it on a page built to explain rather than to sell.
MetriFi publishes these results regardless of outcome, including the experiments that lost, and will keep updating the figures as more experiments finish. If you want to know how your credit union shows up in AI answers today, I am happy to help you start measuring.
Derik Krauss is the Co-Founder of MetriFi (metrifi.com ). He didn't set out to disrupt the status quo; his heart just led him to seek the truth. That passion has fueled 150+ A/B tests and 90+ GEO experiments and helped credit unions generate over $201 million in new loans and deposits. Now, he's building MetriFi to harness analytics and AI for measurable growth. He's a family man who loves Jesus, liberty, basketball, fly fishing, and delicious food. He can be reached at dk@metrifi.com
