For two years the keyword list I'd built behind our cross-border payments content did its job. The educational posts ranked, the rankings fed signups, and the product line they supported, the line I owned end to end, grew from $2M to $76M in annual revenue in its first scaled year. Then, over the spring of 2025, the informational half of that list quietly stopped paying rent. Positions held. Impressions held. Clicks fell month after month on the exact pages that used to be our most dependable earners.
I spent a week blaming tracking. The real cause was sitting at the top of the SERP: an AI Overview answering our keywords before anyone scrolled down to us. This is the post-mortem of how I rebuilt the list once I'd accepted that, and why keyword research in fintech is now a double filter. First you discount every keyword's raw volume by how much of it is no longer clickable. Then you run the survivors past compliance, which kills a different set of keywords for reasons no SEO tool will ever surface. Almost nothing that headlined the old list survived both.
Naming the damage
The pattern wasn't unique to us. Pew Research Center tracked real browsing behavior across roughly 68,000 Google searches and found that when an AI Overview appeared, users clicked a traditional result on 8% of searches, against 15% when there was no Overview. Sources cited inside the summary got clicked 1% of the time. Half the clicks, gone, on any query Google decides to answer itself.
Ahrefs measured the same effect from the other direction: across a study of 300,000 keywords, when an AI Overview appeared, the top-ranking page lost 34.5% of its clicks. The detail that stung was the intent split. 99.2% of keywords that trigger Overviews are informational: “what is a virtual account,” “how does cross-border settlement work.” The exact educational queries our fintech blog was built on. And the floor keeps moving. Similarweb data shows zero-click searches rose from 56% to 69% of all Google searches between May 2024 and May 2025, so the volume column overstates clickable demand by a wider margin every quarter.
Once I accepted all that, the old spreadsheet was unusable. A keyword showing 2,000 searches a month with an Overview parked on it is no longer a 2,000-search keyword. Treated honestly, it might be worth 150 clicks split across everyone on page one.
The rebuilt spreadsheet
The new list is four columns wide where it matters, and two of those columns do all the work.
Raw volume stays, demoted to an input. Next to it: AI Overview present, yes or no, which I checked by hand for our top 150 keywords in a clean browser profile, because the rank trackers were still catching up. Then adjusted expected clicks, which is deliberately crude arithmetic. Informational keywords with an Overview get their expected click-through cut roughly in half, per Pew's 8% versus 15%; transactional keywords keep their normal curve. Last, a compliance risk tier from 1 to 3, which I'll get to.
Crude is fine. The column exists to re-rank, and sorting by adjusted clicks instead of raw volume reordered the entire list. Head terms that had anchored two years of our content calendar dropped below operational queries I used to cut at the first pass.
A keyword with an AI Overview parked on it is a different keyword from the one your tool reports. Price it that way before you brief a writer.
The second filter kills different keywords
Filter one is math. Filter two is legal, and it removes keywords the math filter loves.
If your fintech touches brokerage or investment products, the second filter has a name. FINRA Rule 2210 treats any electronic communication reaching more than 25 retail investors in 30 days as a retail communication, which means principal pre-approval and recordkeeping, and blog posts built from your keyword list count. FINRA fined one broker-dealer $850,000 in 2024 over a social-media influencer program whose posts weren't fair and balanced. We were a payments company, so our review ran through partner-bank and money-transmitter rules instead, but the mechanic was identical: I tiered every planned page before a writer touched it, and compliance owned the verdicts. Tier 1 ships on the normal cycle. Tier 2 needs draft review. Tier 3 means the page probably cannot be written compliantly at all.
Here's the keyword the filter killed. “Guaranteed exchange rate” looked great in the new spreadsheet: real volume, transactional intent, no AI Overview in sight. Our compliance reviewer rejected it in one sentence: you can't put “guaranteed” next to a rate we don't control, and by the time the page carries the qualifiers that make it accurate, the headline is false. I tried “locked-in exchange rate” as a substitute and dropped that too. The person searching “guaranteed” wants that word, and the SERP was owned by providers willing to say it. Tier 3, killed. Two hours of legal review saved us a page that could never rank for its own promise.
The cluster that beat both filters
What rose to the top of the re-sorted list was a cluster I'd been half-apologizing for since the first version: operational questions in the buyer's own words. “Is KYC required for overseas payouts.” “How long does settlement take to China.” “Can a US business receive EUR without an EU entity.” Individually most showed 10 to 50 searches a month; a few showed zero. Ahrefs' analysis of about 4 billion keywords found that 94.74% of them get 10 or fewer monthly searches, which is the statistical way of saying a list filtered at volume > 100 discards most of the demand that exists.
The cluster dodged filter one because these questions are specific and the stakes are money in flight. A generic AI answer doesn't settle “will my supplier get paid before my production slot closes” for the person actually wiring the funds; they click through and verify. It sailed through filter two at tier 1, because the answers are documented facts about process rather than promises about outcomes. Nothing to disclaim.
So I briefed the pages plainly and our writers built them. One question, one direct answer at around 1,500 words, structured after I pulled the top 10–30 ranking pages to see what Google believed the searcher wanted. The drafting workflow I set up scores each page against the brief's question list, 0 to 100: the 40–50 range at outline, the 70s and 80s at full draft. The traffic reports on this cluster still look modest. The sales team's behavior says more: they send these pages mid-deal, and a buyer who lands on one arrives past curiosity, holding a specific fear about a specific wire.
What I'd tell a peer on Monday
Add the two columns before you add a single keyword. Take your current list, run the top 50 by hand, mark which ones carry an AI Overview, halve the expected clicks on the informational ones, and re-sort. That one hour will reorder your quarter.
Then book 30 minutes with whoever reviews your marketing, whether that's in-house legal, a compliance officer, or the partner bank's marketing contact, and tier the survivors before anyone writes a brief. Keywords legal kills late are expensive. It costs a lot less to find out before a writer starts.
Last, mine the long tail where it actually lives. I pulled ours from support transcripts and sales-call notes, phrases customers used that appeared nowhere on our site. I sort by the volume column last now. It counts searches, not clicks I can win.