What 1,385 Financial Homepages Say, in 55 Languages

Test Lab

I made a computer open 3,446 bank websites and write down the first thing each one said. This took a day, produced 1,385 usable headlines in 55 languages, and answered a question nobody asked me: does anyone in finance actually talk to their customer? Norwegian banks do, with emoji. Polish banks never once said the word “you”. The AI layer above the homepage answers before visitors arrive.

Victor Roodhorst24 July 202626 min read

Let me be honest about how this started. Every financial firm I have worked with or pitched to describes itself as client-focused. It is in the strategy deck, the tone-of-voice document, the values page. I have written the sentence myself. And I had a nagging suspicion that almost none of it survives contact with the actual homepage.

Normally that stays a suspicion, because checking it properly would mean reading a few thousand bank websites, and nobody has that kind of time or that kind of problem.

Then I realised I could make a computer do it. So I did, on a Thursday, and by the evening my laptop had quietly opened 3,446 financial homepages around the world, taken down the first thing each one said, and filed it in a spreadsheet. My wife asked what I was doing. "Reading every bank in Norway" is not an answer that improves with elaboration.

What came back is 1,385 usable headlines across 55 languages and more than 60 countries, and it is genuinely stranger than I expected. There is a Turkish insurer whose entire homepage headline is a phone number. There is an American wealth manager that opens with a metaphor about soccer. There is a Norwegian savings bank leading with a warning about jet fuel prices. And there is one former bank domain now serving Indonesian gambling spam, which I had to write a filter to exclude, because otherwise it would have entered the dataset as that bank's brand positioning.

This piece is the full write-up: the question, the method, what it includes and deliberately excludes, the six findings, the two claims I had to throw away when the data disagreed with me, and a hall of fame for the headlines that made me put my coffee down. The dataset and the scripts are at the end, because the entire point of a study like this is that you can check my working.

The question, and why the obvious version of it is wrong

I started with a hypothesis that felt safe. Financial marketing is famously homogeneous, so a large sample of homepages should show a handful of phrases repeating endlessly. Navigating uncertainty. Your partner in growth. Building the future together. Anyone who has sat through a positioning workshop can recite them.

I wrote the first version of the study to count exactly those phrases. It found almost nothing, and that non-result turned out to be the most useful thing that happened, because it forced a better method and a better question.

Here is the problem with hypothesis-first counting. If you decide in advance which phrases are the tired ones and then go looking for them, you will find whatever you brought with you, and you will report a number that reflects your own assumptions rather than the industry's behaviour. The first run scored a set of phrases I had chosen. Almost none of them appeared. The words that did repeat were ones I had not thought to include.

So the method changed. The measurement now runs in two passes and the order matters:

1. Empirical pass first. For each language, count how many firms share each word and phrase. Let the data nominate its own repeated language before I apply any judgement.
2. Coded pass second. Only then score the text against named themes, and only in languages where I have a validated wordlist.

That ordering is the single most important design decision in the study. It is also transferable: if you are auditing your own category's messaging, run the empirical pass before you write the list of words you expect to find.

How the pipeline works, stage by stage

Five stages, each a small script, each one re-runnable on its own. I am describing them in enough detail that you could rebuild this for your own market without talking to me.

Stage one: sourcing a domain universe

The hardest part of a study like this is not the analysis. It is getting a defensible list of who to study. Hand-curating a list means the sample reflects the names you happen to know, which for me would have meant large European asset managers and very little else.

The source I settled on is Wikidata, which is free, global, requires no authentication, and holds an official website property for a large number of companies. The query anchors on the transitive subclass tree of "financial institution", so it captures banks, insurers, asset managers, wealth managers, brokers, pension funds and the long tail rather than only the categories I would have thought to name.

That returned 5,360 unique financial domains across more than 60 countries, once deduplicated by registrable domain.

Stage two: filtering to firms that actually market themselves

The raw 5,360 is not a marketing sample. It contained 177 central banks, 452 cooperative banks and 268 German Sparkassen. A central bank is not competing for retail attention. A village Sparkasse runs a franchise template shared with hundreds of others, so counting them individually would let one template masquerade as an industry-wide pattern.

Excluding those categories left 3,487 firms that genuinely compete on positioning. This is a judgement call and I have documented it rather than buried it, because it is the kind of decision that quietly determines a result.

Stage three: rendering, not fetching

This is the stage that separates this study from something you could do with a simple crawler, and it is where the first version of the analysis went wrong.

If you fetch a financial homepage with a plain HTTP request, you get the raw HTML. On most modern financial sites the raw HTML does not contain the hero headline, because that copy is injected by JavaScript after the page loads. What you get instead is the meta description, which is a different genre of writing: factual, search-oriented, usually some variant of "X is a leading global provider of Y".

Reading meta descriptions and calling them homepage messaging would have produced a confident, publishable, wrong study. So the pipeline drives a real headless browser, waits for the hero to settle, and reads the rendered page.

Stage four: defining "the hero" so a machine can find it

"The hero tagline" is obvious to a human and meaningless to a script, so it needs a deterministic definition. Mine scores every visible text element that intersects the first viewport, using font size multiplied by the square root of its area, weighted slightly toward elements higher on the page and toward heavier font weights. The highest-scoring block is the hero.

Cookie banners, navigation, and consent text are filtered by pattern before scoring, and anything under 15px or outside the first viewport is ignored. The rule is crude but it is consistent, which is what matters: every firm is measured the same way, and you can read the rule and decide whether you trust it.

Stage five: language detection and coding

Language comes from the page's own lang attribute where present, with a diacritic and stopword fallback where it is missing. Then each hero is coded on three language-independent measures, each backed by a per-language wordlist:

• Direct address: does the copy say "you" (or the local equivalent, including formal and informal forms)?
• Concrete product: does it name something purchasable, such as a mortgage, card, account, or insurance policy?
• Abstraction: does it invoke future, trust, growth, partnership, or similar?

These three survive translation in a way that a phrase list never does, which is what makes a 55-language comparison possible at all.

Everything that broke, in order

The clean five-stage description above is how the pipeline works now. It is not how the day went. I am including the failures because every one of them was capable of producing a confident, publishable, completely wrong number, and because the failure modes are more transferable than the findings.

Break one: I guessed the answer and my code agreed with me

Covered above, but it belongs in this list. The first version searched for phrases I had written down from memory. It found nearly nothing, and the natural next move, the dangerous one, is to loosen the matching until the numbers look publishable. That is how you launder an assumption into a statistic. The fix was to invert the order and let the corpus speak first.

Break two: I was reading the wrong text entirely

The first working version fetched raw HTML and read meta descriptions, which produced a tidy result showing that financial firms describe themselves as "leading" and "global". It was internally consistent, it had a decent sample, and it was measuring the wrong artefact. Meta descriptions are written for search engines. Hero copy is written for humans. Publishing the first as the second would have been wrong in a way no reader could have detected.

Break three: my internet dropped and the data did not notice

This is the one that would have been fatal.

Partway through the first full render, the connection dropped. The browser did what browsers do and returned an immediate network error for every subsequent request. The renderer dutifully recorded roughly 2,400 perfectly healthy financial institutions as failures and kept going, finishing in record time with a 38% success rate.

The genuinely dangerous part was the feature I had added for convenience. The renderer skips any domain it has already processed, so a re-run would have treated those 2,400 fabricated failures as settled and never revisited them. My data would have been permanently missing a quarter of the sample, in a pattern determined entirely by when my router hiccupped, and nothing in the output would have looked wrong.

Two fixes. A guard that watches for consecutive network-side errors and aborts the run rather than burning through the queue, and a cleanup pass that distinguishes errors caused by my infrastructure from errors that are real findings about their site. A domain that does not resolve is data. A domain I could not reach because my wifi died is not.

Break four: a second copy of the renderer I did not start

Midway through I found a render process running that I had not launched, competing for bandwidth and writing files. I still cannot fully account for it. What I could do was kill everything, delete every record whose provenance I could not explain, and re-run from a known state.

That cost about an hour and it was not a close call. The alternative was publishing numbers built partly on files I could not explain, which is the kind of thing that stays true right up until somebody asks.

Break five: the corpus contained sentences no bank ever wrote

Wikidata records institutions historically, so some of its URLs now point somewhere else entirely. Once the extraction was working, the English word list came back with cheap car, real estate, affordable coverage and, memorably, slot gacor, which is Indonesian gambling spam.

Dead financial domains get bought. Some become insurance comparison sites, some become casinos. Without a filter, an Indonesian gambling operator's landing page would have been counted as a defunct bank's brand positioning. The junk filter now catches bot walls, error pages, developer-host parking, domains for sale and server defaults: 81 pages excluded in the final run, each logged by category rather than silently dropped.

Five breaks, and four of them would have produced a study that looked fine. That is the actual lesson of building anything data-driven as a marketer: the output almost always looks plausible, so plausibility is worth nothing as a check.

What the study includes, and what it deliberately leaves out

Most data studies bury this section or skip it. It belongs near the front, because the exclusions shape every number that follows.

Included

• Banks, insurers and reinsurers, asset and wealth managers, private and investment banks, brokers and trading platforms, pension funds, and other financial firms.
• Any country. The final set spans the United States (287 firms), Norway (96), the United Kingdom (84), Germany (73), the Netherlands (53), Russia (49), Ukraine (40), Australia (37), Nigeria and the Czech Republic (29 each), Switzerland (27), Canada (24), Poland (23) and dozens more.
• Whatever language the site served to a browser configured for English in a European location. That is an important caveat and I return to it.

Excluded, and why

• Central banks, cooperative banks, credit unions and Sparkassen. Not competing for positioning, or running shared templates that would inflate a pattern.
• 1,324 domains that no longer resolve. Reported as dead, never as a finding about their messaging.
• 81 pages that rendered but were not homepages: 37 bot walls, 34 error pages, 7 developer-host parking pages, 3 domains for sale. One former bank domain now hosts gambling spam. Counting that as a bank's positioning would be indefensible.
• 656 pages that rendered without a detectable hero. Usually heavy JavaScript apps or login walls.

The exclusion that matters most

The coded scoring for named themes runs only on the 685 English-language firms, because that is the only language where I have a validated theme lexicon. Every percentage in the segment analysis below describes those 685, not all 1,385. The cross-language findings use the three translation-resistant measures instead, which is why they can cover the full sample.

I am stating this loudly because it is the single easiest way for a reader, or a journalist, to misread the study.

Finding one: there is no phrase the industry hides behind

The premise I started with was wrong, and it was wrong by a wide margin.

Across the 685 English-language firms, only 21% used any coded generic theme at all. Just 1% used three or more. The remaining 79% said something specific enough that none of my themes caught it.

Look at the bottom of that chart. "Navigating", in any form, appears in three firms out of 685. "Roots" and "goals" appear in two each. The phrases that dominate the industry's self-parody barely exist in its actual homepage copy.

I think there are two reasons this feels wrong to anyone who works in the sector. The first is sampling: we encounter the same forty large brands repeatedly, and large brands are more likely to write in abstractions. The second is that we remember the bad examples. A homepage that says "Fixed rate 4.2% until 2028" does not lodge in memory. One that says "Navigating tomorrow, together" does, and then stands in for the whole category.

The honest version of the industry critique is narrower and more interesting: generic language is not universal, it is concentrated, and it concentrates in a predictable place.

Finding two: the vaguer the product, the vaguer the words

Splitting the English-language firms by segment produces the clearest structural result in the study.

Private and investment banks used generic language at 47.8%, against 16.3% for retail and commercial banks. Insurers sat at 27.5%, asset and wealth managers at 23.4%. The association between segment and generic language is statistically significant: a chi-square across all six segments returns 19.1 with p = 0.0018, and a Fisher exact test comparing private banks with retail banks returns p = 0.0008.

The pattern has an obvious reading. A retail bank sells things with numbers attached. A mortgage has a rate, an account has a fee, a card has a limit. That gives the copywriter something to say. A private bank sells discretion, access and relationship, none of which have a number, and the homepage fills the gap with language about trust and partnership because there is nothing concrete to put there.

Put more bluntly: generic language is what a homepage does when the product cannot be described. That is a positioning problem wearing a copywriting costume, and it is the same failure I have written about in distribution operating models, where the message gets asked to compensate for structure that was never built.

Why I refuse to write "three times more"

The point estimate is 2.9x. It would make a great headline. I am not using it, and the reason is visible in the chart above.

The private bank figure rests on 23 firms. Its 95% confidence interval runs from 29.2% to 67.0%, a spread of nearly 38 points. The retail bank figure, built on 312 firms, has an interval only 8 points wide. So the true ratio could be roughly 1.8x or roughly 4x, and the data cannot distinguish between them.

What survives is the direction, and the direction is solid: the intervals do not overlap and the significance tests are clear. Private and investment banks lead with generic language markedly more often than retail banks. That is the claim. The multiple is not.

I am labouring this because a number like "3x" is exactly what gets pulled into a LinkedIn carousel and repeated until it becomes a fact. If you take one methodological habit from this piece, take this one: publish the interval next to the estimate, and let the width of it discipline your headline.

Finding three: whether a firm talks to you depends on its language

This is the result I did not expect and the one I keep thinking about.

Using the direct-address measure across all 1,385 firms, the spread by language is enormous. Norwegian financial homepages address the reader as "you" in 36% of cases. French in 34%. German 18%, English 17%. Then it falls off a cliff: Ukrainian 6%, Russian 2%, and none of the 22 Polish sites in the sample.

Every one of these firms would say it puts the customer first. Roughly one in three Norwegian and French firms writes that way on the page a customer actually lands on. In the Slavic-language sample it is close to absent.

Some of this is grammatical. Languages differ in how naturally a formal second person sits in marketing copy, and the formal-informal distinction carries social weight in some cultures that English does not have to negotiate. But grammar does not explain a gap between 36% and 0%, because every language in this sample can address a reader directly. German manages 18% with a formal register that is at least as fussy as Polish.

The operator reading is this. If you run a multi-market financial brand, your tone-of-voice document almost certainly says "we speak directly to our customers." That instruction is being followed in some of your markets and quietly ignored in others, and nobody has told you, because the local team is writing in the idiom their market's category leader uses. Category convention is beating your brand guideline, market by market. This is the same failure mode I described in localise, translate, or rebuild: the words get translated, the intent does not.

Finding four: English is more concrete than German or Italian

Here is a result that runs directly against a stereotype I held.

English-language homepages invoked an abstraction in 7% of cases. German managed 13% and Italian 14%. Meanwhile English named a concrete product in 15% of cases, nearly twice the German rate of 8%.

German and Italian are the only two languages in the sample where abstraction outruns concrete product. German financial homepages lead with Vertrauen (trust), Leben (life) and morgen (tomorrow). Czech ones lead with pojištění (insurance).

The stereotype of vague English-language marketing speak simply does not hold here. If anything, the Anglophone habit of putting a rate or a product name in the headline makes English homepages the more concrete end of the distribution. I came into European financial services from an English-language performance background where every headline carried a number, and I had quietly assumed that was the industry's worst instinct. On this evidence it was one of its better ones.

Finding five: some markets run a price war on the homepage

The measure here is whether the hero contains a price, a percentage, a discount or an explicit offer. The spread is wide.

Language	Hero contains a price or offer	Firms
Polish	18%	22
Norwegian	17%	12
Russian	14%	42
Italian	14%	14
French	8%	38
Spanish	7%	46
English	5%	685
German	3%	102

Alfa-Bank's homepage led with "3000 ₽ to everyone who buys OSAGO", a straight cash incentive on a compulsory motor insurance product. Allianz's Czech insurance arm led with "Travel insurance in a few minutes. Now with 35% off." These are not brand statements. They are promotions, sitting in the most valuable piece of real estate the company owns.

Meanwhile German homepages ran promotional copy 3% of the time. Same industry, same regulatory neighbourhood, entirely different idea of what a homepage is for.

Neither approach is correct in the abstract. A promotional homepage converts and teaches the market to shop on price. A brand homepage builds something slower and risks saying nothing. What should worry a multi-market operator is that this choice is usually not being made deliberately at group level. It is being inherited from whatever the local category leader does.

Finding six: Norwegian banks are the strangest and best cohort in the dataset

Norway contributed 96 firms, and by several measures they behave unlike anyone else.

They use emoji. 15% of Norwegian financial homepages carry an emoji in the hero: a house next to a mortgage offer, a green heart on a savings bank. Zero English-language firms did this. Zero German ones. If you had asked me to guess which national banking sector was putting emoji on its homepage, Norway would not have been my answer.

They write the longest headlines. Norwegian heroes average 9.5 words, against 4.3 for Dutch and 5.1 for German. Several Sparebanks run what is effectively a full paragraph, one of them opening with a note about Middle East instability pushing up jet fuel prices.

They attack each other directly. Six of the 72 Norwegian-language firms lead with a variant of "Flytt boliglånet ditt til oss", meaning "move your mortgage to us". That is open competitive poaching in the hero slot, in a category where most markets are still saying "your partner for life".

And they say "you" more than anyone. 36%, the highest in the sample.

Put those together and the Norwegian pattern is coherent rather than eccentric: long, direct, specific, addressed to one person, and willing to name the competitor's product. Sparebank appears in 17% of their headlines, so local identity is doing real positioning work rather than sitting in a values page. It is the closest thing in this dataset to homepages written by people who expect to be read.

Finding seven: 38% of the banks were not there

This one arrived by accident and I nearly filed it as a bug.

Of 3,446 domains attempted, 1,324 did not resolve at all. No server, no redirect, nothing. For a while I assumed my renderer was misconfigured, because a 38% failure rate on a list of financial institutions seemed absurd. It was not misconfigured. Those banks are gone.

Wikidata records institutions as historical entities, so it keeps a bank after the bank stops existing. Retail banking has consolidated ferociously: acquired, merged, absorbed into a parent brand, wound up. The failure rate was concentrated in exactly the places you would expect, among small regional and cooperative institutions, and it was far lower among asset managers and insurers.

So the dead domains are not noise. They are a rough census of consolidation, visible in a dataset built to study something else entirely. Roughly two in five financial institutions notable enough to have a Wikidata entry with a website no longer answer at that address.

Two caveats keep this honest. Wikidata's coverage is uneven, so this is not a clean consolidation rate for any specific market. And it introduces survivorship into everything else in this study: my 1,385 firms are, by construction, the ones healthy enough to still run a website. If failing institutions write differently from thriving ones, and I would bet they do, this study cannot see it.

I am flagging it as a finding rather than burying it in the limitations because it is genuinely interesting, and because a study that only reports what it set out to measure is leaving evidence on the floor.

The words themselves, translated

The empirical pass surfaces the words that actually repeat within each language. Here are the ones that recur most, with what they mean and how concentrated they are.

Word	Language	Meaning	Share of that language
pojištění	Czech	insurance	28%
sparebank	Norwegian	savings bank	17%
boliglån	Norwegian	mortgage	9%
privatbank	German	private bank	9%
crédito	Portuguese	credit	11%
vertrauen	German	trust	recurring
leben	German	life	recurring
morgen	German	tomorrow	recurring
budoucnost	Czech	future	recurring
вклад	Russian	deposit	recurring
страхування	Ukrainian	insurance	recurring

Read the left column top to bottom. The most-repeated words in Czech, Norwegian and Portuguese finance are all products. The most-repeated words in German finance are trust, life and tomorrow. That single contrast is the study in miniature.

What the taglines actually look like

Numbers flatten texture, so here is a sample verbatim, as served on 24 July 2026:

Aasen Sparebank (Norwegian): "Vi er lokalbanken - i gode og andre dager." We are the local bank, in good days and other ones.

AlpenBank (German): "Wir für Sie. Schön, dass Sie bei uns sind." We, for you. Good that you are with us.

Allianz pojišťovna (Czech): "Cestovní pojištění za pár minut. Nyní s 35% slevou." Travel insurance in a few minutes. Now 35% off.

Bank 131 (Russian): "Банк цифровой экономики." Bank of the digital economy.

Jupiter (English): "The value of active minds."

1789 Capital (English): "Funding the Next Chapter of American Exceptionalism."

First State Bank Nebraska (English): "Earn up to 6.05% APY with a First Rate Checking Account."

Those last two are worth sitting with. They are both English, both American, both financial, and they are doing entirely different jobs. One is a political identity statement. The other is a rate. The distance between them is larger than the distance between most national averages in this study, which is a useful corrective to any argument that culture determines everything.

The hall of fame

Averages hide the good stuff. Somewhere in 1,385 headlines are the ones that made me stop and read them twice, and it would be a waste of a dataset not to show you. All verbatim, all as served on 24 July 2026.

The ones that forgot to finish

Provident Mutual Life Insurance leads with "Welcome to". That is the whole headline. Welcome to what remains unresolved.

9 PSB Ltd goes with "financial". One word. Lowercase. A&D Mortgage counters with "partners", also one word, also lowercase, and between them they have accidentally written the shortest possible parody of the entire industry.

Al Akhdar Bank's homepage headline is "Lien innacessible", which is French for "link inaccessible". A broken link, rendered at 48 pixels, sitting where the value proposition goes. I checked twice because I assumed my scraper had failed. My scraper was fine.

The ones that lead with a number, but the wrong one

Ankara Sigorta's entire hero is "444 1936". A phone number. No verb, no product, no promise. In fairness this is the most functional headline in the dataset, and if the goal is "call us", it is undefeated.

Artisan Partners opens with "$183.4 Billion", which is assets under management presented as a personality. Banco de La Pampa goes further and leads with its actual fee schedule: free savings account maintenance, annual renewal service commission $185,130.00, other banks' network fee $2,783.00. It is a homepage that opens with a price list. I have complained for years that financial marketing hides its pricing. I would like to formally withdraw the complaint in this one case.

The one with the soccer metaphor

Merrill, a wealth manager owned by Bank of America, opens with this:

"In soccer, trusting your team is a key ingredient of success. At Merrill, it's no different, as our experienced advisors can assist you with personalised..."

An American firm, using the word "soccer", to explain trust, to people with money. Every clause of that sentence is a decision somebody defended in a meeting.

The Norwegians, again

Bjugn Sparebank's homepage on the day I ran this opened with a bulletin about Middle East instability pushing up jet fuel prices and possibly cancelling flights, with an explainer on what you should do about it. This is a local savings bank operating, unprompted, as a travel news desk.

SpareBank 1 Østlandet and SpareBank 1 Helgeland both lead with "Ror du deg unna? ⚽ 🚣", roughly "are you rowing away?", accompanied by a football and a rowboat. I have read the page and I am still not certain what is being sold. I am, however, certain that I would open the app.

And my own broker turns up: DEGIRO runs "Financial power to you 💪" in English, "ve vašich rukou 💪" in Czech, "tehokkuutta 💪" in Finnish. Same bicep, four markets. Say what you like, it is at least a decision, applied consistently, which is more than most of this dataset manages.

The one that has been doing this for 230 years

Bankhaus Seeliger, Germany: "Geld und Vermögen sind Vertrauenssache. Und Vertrauen, das wächst mit der Zeit." Money and wealth are a matter of trust, and trust grows with time. Then it notes it has been serving customers for 230 years.

By every measure in this study that headline is textbook generic language: abstraction, no product, no "you". It is also completely earned, because they can prove the claim, which is exactly the distinction the coding cannot make and a human reader can. Keep that in mind for every percentage in this piece.

What I got wrong

Two claims died during this study. Both are worth recording, because a data study that reports only its successes is a marketing asset rather than research.

The premise itself

I set out to prove that financial homepages all say the same thing. They do not. The first version of the analysis scored a lexicon of phrases I had confidently written down in advance, and found almost nothing, because I had guessed wrong about which words repeat. Had I published that first pass with a headline like "38 of 100 firms say the same thing", the first person to check the data would have found the claim unsupported.

The fix was the two-pass method described earlier. The lesson is older than this study: if your analysis only counts what you expected, it can only confirm you.

The Dutch investor-relations claim

Partway through I noticed that Dutch financial homepages seemed to lead with corporate and investor-relations content rather than customer messaging. ABN AMRO's hero was "Cijfers eerste kwartaal 2026", first-quarter figures. It made a satisfying story about Dutch financial culture.

Then I counted. The pattern rested on one firm out of 22. It is not a finding, it is an anecdote with a percentage attached, and it is not in the results.

The country story that isn't there

I also went looking for national differences in generic language and found nothing usable. The United States sits at 24.9%, the United Kingdom 20.8%, Nigeria 22.2%, Australia 11.4%. Every confidence interval overlaps every other. I would have enjoyed writing that section. There is no evidence for it, so there is no section.

What to actually do with this, by seat

A study that ends at "interesting" has wasted your time. Here is what I would do with these findings depending on where you sit.

If you run marketing at a private bank or wealth manager

You are in the segment with the highest generic-language rate in the study, and the reason is structural rather than creative. Your product genuinely is harder to put a number on. That is not an excuse, it is a diagnosis, and it means the fix is not a better copywriter.

The firms in this segment that escaped the pattern did it by naming something specific: a minimum, a coverage model, a named specialism, a proof point with a date on it. Bankhaus Seeliger gets away with pure abstraction because it can say 230 years. If you cannot point at an equivalent, your homepage will keep reaching for trust and partnership, because there is nothing else in the drawer.

The question to take into your next positioning session is not "how do we sound less generic". It is "what can we say that a competitor could not put their logo on", and if the honest answer is nothing, that is a product and proof problem that marketing has been asked to paper over.

If you run a multi-market brand

This is where the language finding bites. Your tone-of-voice document says you speak directly to customers. In this dataset, whether that instruction is followed varies from 36% to 0% depending purely on the language of the site.

What is almost certainly happening is that each local team writes in the idiom of their market's category leader, because that is what reads as credible locally, and nobody at group level has ever compared the outputs side by side. Category convention beats brand guideline, quietly, market by market.

The cheap diagnostic: screenshot the first viewport of every market site you run, put them on one slide, and translate each headline literally into English. You will see the drift in about ninety seconds. It is the same exercise I described in localise, translate, or rebuild, and it is uncomfortable in a productive way.

The decision that follows is a real one and I am not going to pretend it is obvious. Enforcing group voice everywhere buys consistency and costs local credibility. Letting each market follow local convention buys relevance and costs any coherent brand. What you should not do is what most groups do, which is write the guideline, never check, and believe the deck.

If you run a retail bank or an insurer

You are already the most concrete segment in the study, and the pressure you will face is to become less so. Someone will eventually propose replacing the rate on your homepage with a statement about your purpose. The data suggests that would move you toward the segment with the least differentiated messaging in financial services, in exchange for nothing measurable.

The Norwegian cohort is the model worth stealing here: long, direct, specific, addressed to one person, and willing to name a competitor's product. "Move your mortgage to us" is not sophisticated. It is legible, and legibility is scarce.

If you are considering running your own study like this

The barrier is much lower than it looks. This entire pipeline is five short scripts, a free public dataset, and a headless browser. It cost nothing but a day. What it produces is something no competitor can copy on the day you publish, because they would have to build it too, and it makes you the primary source rather than the ninth summary of somebody else's report.

The bar is not technical sophistication. It is the discipline described in the failures section: measure what you did not expect, publish your exclusions, and let the confidence interval decide how loud your headline gets.

Limitations you should hold against this

A study is only as trustworthy as the list of things it cannot tell you.

1. The coded themes are English-only. Those percentages describe 685 firms, not 1,385. The cross-language findings use different, translation-resistant measures. Do not mix the two.
2. Wikidata is curated, not random. It over-represents notable firms and under-represents boutique wealth managers, which are exactly the firms most likely to write vaguely. If anything this biases the generic-language rate downward.
3. Language served depended on my location. The browser ran from a European IP with English locale headers. Some global firms geo-served a local-language page anyway, which is itself informative but means "site language" is not always "the firm's primary language".
4. The hero definition is a heuristic. Font size times root area is defensible and consistent, but on some layouts it will select a promotional banner over the brand statement below it.
5. Survivorship. 1,324 dead domains means the sample skews toward firms healthy enough to still run a website.
6. One day's snapshot. Homepages change constantly. Several of these firms are running a campaign that will be gone next month.
7. Small cells. Private banks (23), brokers (22), Italian (14), Turkish (13). Directional only. The confidence intervals are published for exactly this reason.

How to run this on your own market

The pipeline is five scripts and none of them are clever. If you want to audit your own category, this is the sequence.

1. Build the universe from a public source, not from memory. Wikidata's SPARQL endpoint is free and needs no key. Anchor on the subclass tree for your category so you catch the firms you would not have listed.
2. Write down your exclusions before you look at results. Which entities are not competing for attention? Which run shared templates? Decide, document, move on.
3. Render, do not fetch. If you read raw HTML you are reading meta descriptions and calling them positioning.
4. Define your extraction rule mechanically and apply it identically to everyone, including the firms you like.
5. Run the empirical pass first. Let the corpus nominate its own repeated language before you write a single lexicon entry.
6. Compute confidence intervals for every rate, and let the widest interval decide how strong your headline is allowed to be.
7. Publish the exclusions and the failures. They are the reason anyone should believe the rest.

The query that builds the universe

This is the whole sourcing step. It runs against Wikidata's public SPARQL endpoint, needs no account, and returns every entity that is an instance of anything descending from "financial institution" and has an official website recorded.

SELECT ?companyLabel ?website ?countryLabel ?typeLabel WHERE {
  ?type wdt:P279* wd:Q650241 .
  ?company wdt:P31 ?type ; wdt:P856 ?website .
  OPTIONAL { ?company wdt:P17 ?country. }
  SERVICE wikibase:label { bd:serviceParam wikibase:language "en". }
}

The important character in there is the asterisk in wdt:P279*. It means "subclass of, transitively, to any depth", so instead of listing the categories I happen to think of, it walks the whole tree and returns savings banks, takaful operators, credit institutions and sixty other types I would never have enumerated. Paginate it with LIMIT and OFFSET and the full result set arrives in about a minute.

Swap Q650241 for your own category's Wikidata identifier and the same query builds your universe. That substitution is the entire difference between this being a financial services study and being a study of anything else.

The rule that defines a headline

The extraction rule matters more than it looks, because it is the operational definition of the thing being studied. Mine, in plain English:

• Consider every visible text element whose box intersects the first viewport at 1366x900.
• Discard anything under 15px, anything shorter than 8 characters or longer than 220, and anything matching the navigation, cookie or consent patterns.
• Score what remains as font size multiplied by the square root of the element's area, multiplied by 1.15 if the font weight is 600 or heavier, and by 1.25 if it sits fully within the first screen.
• Take the highest scorer. That is the headline.

It is a blunt instrument and it gets things wrong. On a site running a campaign banner above the brand statement, it will pick the banner. What it does guarantee is that Merrill and Bjugn Sparebank were measured by exactly the same rule, which is the only property that actually matters in a comparison.

If you build your own version, write your rule down before you look at any results, and resist the urge to tune it once you have seen which firms it flatters.

Two engineering notes

Two engineering notes that cost me time. A dropped network connection silently marked about 2,400 live sites as failures, and because the renderer skips already-processed domains on restart, that would have baked the damage in permanently. The fix was a guard that aborts on consecutive network errors rather than burning through the queue. Separately, a stale domain list means some URLs now resolve to parking pages, bot walls, or in one case gambling spam sitting on a former bank's domain. Filter those explicitly, or your corpus will quietly contain sentences no bank ever wrote.

That verification habit is the same one I described in nine months of building as a marketer: the building is cheap now, and the checking is the entire job.

Questions people asked while I was building this

What do most financial companies actually say on their homepage?

There is no single dominant phrase, which was the finding that killed my original premise. The most common recurring theme in English hero copy was trust, at 4.1% of firms. Every coded theme sat below 5%. What varies far more than the words is the kind of statement: a product, a price, an abstraction, or a direct address to the reader.

Isn't "navigating uncertainty" everywhere?

No. It appears in three firms out of 685 English-language sites. It feels ubiquitous because we all encounter the same few dozen large brands repeatedly, and because a bad headline is far more memorable than a functional one. "Fixed rate 4.2% until 2028" does not lodge in the memory. "Navigating tomorrow, together" does, and then stands in for the whole category.

Why render the pages instead of just downloading the HTML?

Because on most modern financial sites the hero copy is injected by JavaScript after load. A plain HTTP fetch returns the meta description, which is written for search engines rather than humans. Reading those and calling it homepage messaging produces a study that is coherent, plausible and measuring the wrong thing.

Why only 1,385 firms when you sourced 5,360?

1,873 were excluded as non-marketers (central banks, cooperatives, Sparkassen), 1,324 domains no longer resolve, 656 rendered without a detectable hero, and 81 were bot walls, error pages or parking pages. Every exclusion is a documented rule, and the funnel chart above shows each step.

Can I use these numbers for languages other than English?

Only the cross-language measures: direct address, concrete product, abstraction, and promotional content. Those use per-language wordlists and cover all 1,385 firms. The named theme percentages, including the segment comparison, describe only the 685 English-language firms. Mixing the two is the easiest way to misread this study.

Is the private-bank finding reliable?

The direction is; the size is not. It rests on 23 firms, with a 95% confidence interval running from 29.2% to 67.0%. The significance tests are clear (Fisher exact p = 0.0008, chi-square across segments p = 0.0018), so private and investment banks genuinely do use generic language more than retail banks. But the data cannot tell you whether the gap is roughly double or roughly quadruple, which is why the phrase "three times more" appears nowhere in this piece.

How often will you re-run it?

Annually, with the next run planned for Q1 2027. The pipeline is resumable and the domain list is versioned, so the second run is a more interesting article than the first: it will show what moved.

Victor Roodhorst works in marketing in European asset management and publishes the AI Share of Voice Index for European Asset Management. He came to the industry from B2C performance acquisition, which is why he still expects a headline to contain a number. Reachable at victor@roodhorst.eu.

Score your own first line with the interactive tool: score your own first line.