Generative engine optimization, usually shortened to GEO, is the practice of structuring an institution’s website and published content so that AI assistants such as ChatGPT, Gemini, Perplexity and Google’s AI Overviews name and cite that institution when they answer a question.
That is the whole definition. The rest of this guide is about what it means in practice for a bank or a credit union, what the evidence actually supports, and what it does not.
The short version: when someone asks an AI assistant where to open a certificate of deposit in their town, the assistant returns a handful of named institutions. Those names are an outcome. Something about those institutions’ websites made them retrievable, quotable and relevant enough to be named, and something about everyone else’s did not. GEO is the work of being on the named list.
How this differs from search engine optimization
Traditional SEO competes for a position in a list of links. The reader still chooses, still clicks, and still lands on your page. Ranking fourth is worth less than ranking first, but it is a long way from worthless.
An AI answer works differently. The assistant reads a set of sources and writes a single synthesized response, naming a few institutions inside it. A synthesized answer has no fourth position to occupy. Either the answer names you or it does not, and if it does not, the reader may never learn you were an option. A member comparing savings rates can complete the entire comparison inside one conversation without visiting a single institution’s website.
This changes what the work optimizes for. SEO asks whether a page can rank. GEO asks whether a page can be read, quoted and attributed by a system that is summarizing rather than listing. Those overlap, since a page no crawler can reach will not be cited either. But they are not the same job, and the gap between them is where most institutions currently sit.
The four questions that decide whether you can win an answer
Most GEO advice in circulation collapses into a single instruction: publish more. That is not what our data shows. In our experiments, whether an institution can appear in a given AI answer comes down to four questions, and only one of them is about publishing.
1. Can an assistant reach the information at all? Content that cannot be retrieved cannot be quoted, however good it is. Rate tables that render only inside a script, product terms that live in a PDF, and pages behind an interstitial are all harder to retrieve than plain page text.
2. Do you have content that directly answers the question being asked? Not the topic. The question, in something close to the words a person used to ask it.
3. Is this a question where institutions get named at all? Some questions are answered generically, with no institution named, by anyone. There is no slot to compete for.
4. Do you have a realistic reason to win it? On some questions the answer is decided by your rate rather than your website, and no amount of page structure changes that.
The order matters. The first two are usually fixable with work you can schedule. The second two decide whether that work is worth the money, which is why we score them before writing anything. Most disappointing GEO programs we have seen went straight to question two and never asked three or four.
The rest of this guide works through each in turn.
Why this matters specifically for a bank or credit union website
Three things make financial institutions an unusual case.
The questions are local and product-shaped. People do not ask AI assistants abstract questions about banking. They ask where to find the best CD rates in a specific place, or where to open a high-yield savings account near them. These are exactly the questions an assistant answers by naming institutions, which means the opportunity is real rather than theoretical.
The competitive set depends on the question. This is worth getting right, because it decides where effort is worth spending. On a local, product-specific question, a community institution is often competing with the other institutions in its market, and that is a fight it can win on relevance and specificity. On a superlative price question, such as where to earn the most on savings, the same institution is competing with online-only banks and national comparison sites, and rate is doing most of the work. Same institution, same website, very different odds depending on what was asked.
The content is regulated. Advertising a rate carries disclosure obligations, and that is one reason many institutions keep rates inside a widget or a PDF rather than in the page text. That has a real cost in retrievability, and it is a genuine tradeoff rather than an oversight. It is also often solvable within the disclosure rules, but any change to how a rate appears on a page is an advertising decision and belongs in your normal compliance review rather than in a marketing sprint.
A fourth factor applies to institutions with branches. Questions about where to open an account are partly local-presence questions, and the answer often draws on location data, hours and contact details as much as on product pages. Institutions with many branches frequently have that information duplicated across several inconsistent pages, which gives an assistant conflicting facts to reconcile.
There is also a timing argument, and it deserves to be stated carefully rather than optimistically. In the markets we track, the institutions named most often are usually the largest ones, and that has not changed. What has changed is that a smaller institution can now enter that set. Pearl Hawaii did exactly that, reaching the named list on the questions its work targeted without displacing the biggest names in Hawaii. That is a real opening. It is not a promise that size stops mattering.
What AI assistants actually cite
This is the part most worth understanding, because it has been measured rather than guessed at.
Our co-founder Derik Krauss analyzed 7,500 AI search responses to find what the pages being cited had in common. The pattern was consistent, and it has held up since: AI tends to reference detailed, objective web pages that are highly relevant to a user’s prompt.
That is an association rather than a proven mechanism. The analysis shows what cited pages tended to look like, not what caused a model to choose them. We treat it as a working rule because building pages against it has repeatedly moved the number in controlled experiments, which is the strongest evidence available from outside a model. Read that way, each of the three words is doing work.
Detailed. Thin pages tend to lose. An assistant summarizing an answer needs something substantial enough to draw a specific sentence from.
Objective. Marketing language is hard to quote. Facts are easy to quote. A page that states terms, numbers and conditions plainly gives an assistant something it can repeat without editorializing.
Highly relevant to the prompt. This is the one most institutions miss. The page has to answer the question somebody actually asked, in something close to the words they asked it. A general savings page is not a relevant answer to “where in Kona can I find the best credit cards”.
That helps explain something that otherwise looks strange. There is no evidence that assistants work from a private ranking of which institutions are good. They read pages anyone can read, and the ones they quote tend to be the ones that make the relevant fact easiest to lift. So when an institution is missing from an answer, the explanation is usually simpler than a judgment about its quality: nothing on its site answered that particular question in a form worth quoting.
There is a consequence worth spelling out. When your own pages are hard to quote, the answer does not come back blank. It gets built from an aggregator, a rate comparison site or a directory listing, none of which you control and several of which carry stale figures. You can end up described in an AI answer by a page you have never seen, using a rate you no longer offer. Fixing your own pages is partly about being cited and partly about displacing a worse source.
All of which assumes the assistant can reach the page in the first place, which is question one from the list above. In our experience, the retrieval failures described there are not marginal: content locked inside a script, a PDF or an interstitial is frequently missed entirely, and a page that is never read cannot be judged detailed, objective or relevant.
What we have measured
We would rather show results than project them, so what follows is what our own experiments produced, including the parts that did not work.
First, what these percentages mean. Every figure below is a share of tracked AI responses that named the institution, sampled from a set of runs in a measurement window rather than counted across every answer the model could ever give. Samples are small, so the percentages move. In one of our own measurements, a question that read 100% across three responses settled at 56% when we re-ran it across nine. Both numbers were accurate readings of what they sampled; only the second was stable. Treat a single 100% as a direction, not a constant, and treat movement of a few points as noise.
Three credit unions went from nothing to being named consistently, by publishing. Northwest Preferred FCU published a 3,600 word page built around a single question, “Who has the best high yield savings account in Stayton, Oregon?”, and went from 0% to 100% visibility on it. Lone Star Credit Union did the same for “Where in East Texas could I refinance my auto loan?”, also reaching 100%. HFS Federal Credit Union did it for “Where in Kona can I find the best credit cards?”, reaching 100% in the measured week and averaging 88.89%. In each case the page was long, specific, and written to answer one real question directly.
Pearl Hawaii points to retrieval, rather than content volume, as the binding constraint. Visibility across six tracked questions about Oʻahu deposit products moved from zero to 27.38% over three experiments. The instructive part is the sequence. Two rounds of publishing produced only a small foothold, and the large gain came in a third round that published nothing new and instead made existing product pages structured and retrievable. That sequence is the argument. We can show the gain followed the structural round and landed on the questions that round targeted, but the site’s archived history was incomplete, so we could not isolate which individual change carried it, or rule out a contribution from the earlier rounds arriving late.
Zing Credit Union suggests balance can beat advocacy. A long-form article covering business banking options in the Denver Metro area named and fairly described competing institutions alongside Zing. Visibility moved from 66.67% to 85.71%. This is a single-prompt result from an already-high baseline, so we hold it as directional rather than settled: it is consistent with the idea that content reading like an honest comparison is more quotable than content reading like a pitch, and it is not on its own proof of it.
Read together, these are not competing lessons. They point the same way from different directions: cited pages tend to be detailed, objective and relevant, and an institution can fail that test either by not having the page or by having a page nothing can read.
Which questions you can realistically win
This is questions three and four from the framework, and it is where most GEO budgets are won or lost. Not every question is winnable, and knowing which is which saves money.
In our tracked data, the questions that moved carried local, product-specific intent. The ones that did not were either definitional or driven purely by price.
Definitional questions, such as the difference between a certificate and a certificate of deposit, get answered generically. Across every tracked response to that exact question, no institution was named at all, not ours and not a competitor’s. There is no slot to win, so structure and length change nothing.
Price-superlative questions behave differently. When someone asks where to earn the most on savings in a specific place, the answers go to whoever posts the highest headline yield at the time, which is sometimes an online-only bank and often a local institution running a promotional rate. If you are not competing on headline yield, page structure is unlikely to change that answer. That is a pricing reality rather than a content gap.
So the questions worth pursuing are the ones where a local, product-specific answer is the natural one, and where institutions are already being named. Our operating rule is that if nobody is named in the answers to a question today, we do not treat that question as an opportunity yet. That can change as a question matures, which is a reason to re-check rather than to write it off permanently.
Checking this first is cheap. Ask the assistant the question yourself, a few times, and look at whether any institution is named at all, and whether the ones named are competing on rate or on relevance. That tells you which of the four questions you are actually facing before anybody writes a page.
Where to start, and in what order
The order matters more than most teams expect, and getting it backwards wastes money.
First, make sure what you already have can be read and quoted. If your core product pages are invisible to retrieval, publishing more will not fix it, because the new pages usually inherit the same problems. At Pearl Hawaii, the round that did this work, and published nothing else, coincided with a roughly tenfold move. The checklist that round worked through:
• Structured data describing the products, their rates and their terms
• Current rate, minimum deposit and term shown in plain page text rather than only inside a calculator, routed through compliance as with any rate presentation
• Page copy and headings naming the place served, the product, and the current rate
• A short section answering why someone would choose this product here, with a linked set of frequently asked questions
• Plain local phrasing, using the terms people actually type rather than internal product names
• Branch locations, contact details and a visible last-updated date
• Metadata on the about page, and member testimonials
• Page speed and mobile improvements
We cannot tell you which of these mattered most, for the reason given above. What they have in common is the useful part: almost every item takes a fact the institution already had and moves it somewhere an assistant can read it.
Then publish, deliberately, against real questions. This is where the largest gains in our data have come from, and it is ongoing work rather than a one-time fix. Northwest Preferred, Lone Star and HFS did not publish more content in general. They each published one thorough page aimed at one question they had chosen, and it worked. The more questions you cover this way, the more answers you can appear in.
Then keep going, because positions can decay. Nine months after Pearl Hawaii’s gain, a re-measurement found most of it had persisted and some had eroded. We have seen the same pattern elsewhere, including a tracked question that fell from 80% to 50% between windows. Visibility earned this way behaves more like a position that needs defending than a result you can bank, so budget for maintenance rather than for a project with an end date.
How MetriFi approaches this work
MetriFi provides generative engine optimization for banks and credit unions. We track how often an institution is named in AI-generated answers to the questions its members actually ask, work out which of those questions are winnable, do the work, and then measure whether the number moved.
The measurement is the part we care most about. Every recommendation is run as an experiment with a baseline before the work and a measurement window after it, and we publish the results, including the ones that did not go the way we expected. The Pearl Hawaii and Zing write-ups report decay, small samples and causes we could not attribute, alongside the gains. We would rather hand you a number you can trust than a number that flatters us.
If you want to know where your institution currently stands in AI answers about your own products, you can talk with us, or see what MetriFi GEO tracks and reports.
If you would rather see the method than a description of it, the Pearl Hawaii and Zing write-ups each show a full cycle: the baseline before the work, what changed, the measurement window after it, and what we could and could not attribute.
