Summarize with AI
Yes, say which LLM you use, but not on your homepage. Name your model providers on your trust page, in your docs and in your security answers, where the people who need that fact go looking for it. Keep model names out of the hero, the headline and the logo strip, where they tell every other visitor that your product is someone else's model with your logo on top.
It sounds like a copy decision. It is really about who your website is talking to. A security reviewer needs the provider's name to approve you. Your actual buyer mostly does not care which model runs underneath, and reads the name as a hint about how thin the product is. You know exactly what you built on top of that model. A stranger only knows what the first screen tells them.
A subtitle tool whose hero says GPT-5 before it says subtitles
Picture an AI tool for documentary producers and small post-production houses. This is an example, not a client. An editor uploads a locked 92-minute cut along with the delivery spec from the streaming service that bought the film. The tool returns subtitle files in eleven languages, timed to the cut, plus a check report against that spec. Most lines pass. A handful are flagged in red, each with a timecode. Line 412 runs at 21 characters per second, and the spec allows 17. Line 1,088 crosses a shot change. A contributor's surname is spelled two ways in the French file, Okonkwo and Okonkow.
Those red lines are the whole product. Miss them and the files fail the streamer's quality check, and the delivery date slips while someone fixes them by hand.
Now open the homepage. Above the headline sits a small pill: Now running on GPT-5 and Claude. The headline reads Frontier AI models, built for subtitling. The subhead says Every cut is transcribed and translated by the latest models from OpenAI and Anthropic. Under the Try it free button, where a customer logo strip would usually go, a row headed Built with shows the three model providers' logos. The words delivery spec, timecode and quality check do not appear until the fourth section, under a heading called What you get.
Behind the page sits work that no provider ships. The team built a test set of a few hundred films whose subtitles a professional subtitler had already corrected, and grades every change against it. They wrote each streamer's delivery spec as checks the system runs line by line. In the spring they moved two languages to a different model, because it made fewer errors on that set. The pill on the homepage still names the old pair.
What the founder hears after the demo
In demos, the first question from producers is some version of: so this is ChatGPT translating my film? That line is an illustration, not a quote from a call. He spends the next five minutes on the checks instead of showing the red line at 21 characters per second.
On a seed call, an associate at the fund pastes a page of the sample transcript into a chat app while they talk and gets back a clean French version. The follow-up email asks one question: what does the product do that the model alone does not? He writes a long reply about the corrected test films and the spec checks, and the deal goes quiet for three weeks.
Meanwhile a distributor's content security team sends a form. Unreleased films are the thing they guard hardest. Question twelve asks which AI providers receive footage or transcripts and what they keep. The answer lives in a doc his CTO wrote, not on the site. He has edited the pill three times this year, once per new model release, and the trust page not once.
Who actually asks which model you use
Four people land on an AI company's site, and only two of them want the model name.
The reviewer who can block the deal. At a bigger customer, someone from IT, security or compliance has to approve any tool that handles client files. For them the model provider is a subprocessor, a company that touches customer data, and they cannot sign off without its name, where it runs and what it keeps. Hiding it reads as hoping they would not ask.
The technical evaluator. An engineer at the customer, or your user if you sell to developers. They want versions, limits, whether they can pick the model or bring a key, and what happens when the provider has a bad day. For devtools this person is often the buyer.
The investor. He asks which model, but the question under it is what is left if that model changes or its maker builds your feature. Naming the provider louder answers the surface question and makes the real one harder.
The buyer. The producer never asks. When the name is in front of him anyway, he reads it as a clue that the product is the model, which is the one conclusion you most need him not to reach.
Why the hero is the wrong place, even for a true fact
Most advice comes in two flavours. Hide the model, because naming it looks like a wrapper. Or lead with it, because transparency builds trust and a famous model lends its fame. Both are wrong. Hiding it fails the reviewer, who needs the name in writing. Leading with it fails everyone else, because a stranger decides what kind of company you are in about four seconds, and you spent those seconds on a part anyone can rent.
Here is the claim some founders will push back on: a model name in your hero is a cost even when the model is the best one available. It describes the one layer your competitors can buy on the same day you did. The subtitle tool's hero could be pasted onto any rival that signed up for the same API. The red line at 21 characters per second could not.
The Series A question has changed shape. Investors now probe whether you keep a reason to exist once a big model maker ships something far better. A row of provider logos under your headline answers that before the meeting starts, and the answer it gives is no. Carta puts the Series A bar for AI startups near 3.5 million in ARR in Q1 2026, against roughly one million three years earlier, and seed AI rounds are priced roughly 42 percent above non-AI peers. Being read as a thin layer is expensive at both prices.
There is a margin cost too. AI product builders average around 52 percent gross margin against 70 to 80 percent for traditional software, because every model call sits in cost of goods. Your model is a cost lever you will keep pulling, and a hero that names it turns every switch into a public event.
The Dualite case study shows the same pull in another category. Dualite is an AI app builder, a space crowded with tools that run on much the same models. Its non-technical visitors could not tell it apart from a dozen other builders, so the site was designed to lead with brand and imagery and make clear who Dualite was before explaining what it did.
Where the model name belongs on your site
Put it where the people who need it already look, written plainly enough to copy into a form.
The trust page. Add a section called Which models process your data. List each provider by name, the job it does, the region it runs in, and what it keeps. Put the date you last changed it at the top. Our guide to what a trust page for an AI product should answer covers the rest of that page.
Answer the training question in the same place, because it is often the real question behind which model. Say whether you train on customer data, and what your provider's terms say. OpenAI's API documentation, for example, states that data sent to its API is not used to train its models unless the customer opts in, and that abuse monitoring logs are kept for up to 30 days by default. Quote your own provider's terms and link them, rather than paraphrasing from memory.
The docs. A short Models page for technical readers: which model handles which step, the versions, any limits a user will hit, whether they can choose or bring a key, and what happens on an outage.
The security form. Keep one written answer and paste it everywhere, so the trust page, the docs and the form never disagree.
Saying it without tying yourself to one provider
The sentence that works names the providers and then points at your own work. For the subtitle tool: We use models from OpenAI and Anthropic, chosen for each language. Every model change is tested against films a professional subtitler has corrected before it reaches your account. The first sentence satisfies the reviewer. The second tells the investor and the engineer where the value lives.
That framing also makes switching boring, which is what you want. When you move a step to a new model, do four things. Write a dated changelog entry that says what moved, why, and what your test showed. Update the provider list on the trust page and its date. Email customers whose contracts ask for notice of new subprocessors. Leave the hero alone.
That last step is the test for the whole page. If changing your model would mean editing your headline, your headline is about the wrong thing.
A one-week fix for a page that leads with the model
- Monday. Search your own site for every model and provider name: the product names, the company names, and phrases like powered by and built on. Paste each hit and its page into one list.
- Tuesday. Remove the names from the hero, the headline, the nav, the buttons and any logo strip. Replace the hero line with the job and the result, the way the subtitle tool would say subtitles in eleven languages that pass the streamer's check the first time.
- Wednesday. Write the Which models process your data section for the trust page: providers, jobs, regions, retention, training, and the date.
- Thursday. Write the Models page in your docs for technical readers, and copy the same answer into your security form template.
- Friday. Write the thirty-second answer to what happens if the model changes. It should name your test set, your data and your workflow, not your provider. Then read your homepage next to your deck, as covered in what investors look at on your website, and make sure both tell that story.
If a reviewer can approve you from the trust page and a stranger can say what you do from the hero, the model name is in the right place.
A homepage that leads with the model usually means the page has not found its own noun yet, and that is a legibility problem before it is a copy problem. Studio Maydit designs and builds sites for AI founders in Framer, Webflow and custom code, with website design as fixed scope over three to four weeks or as a monthly retainer. Dualite reached 100,000+ users in seven months after design work that supported a repositioned ICP. If you want a second opinion on what your first screen should say instead of a model name, book a call.
