Summarize with AI
A typical free to paid conversion rate for a freemium SaaS product sits between 2 and 5 percent. Free trials that ask for no card convert around 18 percent, and trials that take a card at signup convert around 49 percent. Those are benchmarks measured on SaaS and product-led companies, not on AI products, and the sources disagree with each other, so treat them as a range and not a target.
If your rate is low, the pricing page is the last place to look. Free-to-paid is capped by activation. A person who never got real value from the free plan has nothing to pay for, whatever the plan table says. For an AI founder that cap matters more than usual, because every free user who leaves after one try was also paid for in inference.
Why 2, 18 and 49 percent are not the same question
The three numbers look like a ladder, as if a good team climbs from freemium to opt-out. They are not a ladder. Each one divides by a different crowd. A freemium rate divides by everyone who ever made a free account, including the curious and the lost. An opt-in trial divides by people who chose a clock. An opt-out trial divides by people who already typed in a card, so most of the sorting happened before the trial started.
That is why moving to a card-required trial almost always lifts the rate and often does little for revenue. You have not converted more people. You have turned away the ones who would not have paid, before they could try the product. Some founders want that trade. It is still a change to the denominator, not a fix.
So the useful question is not whether your rate is good. It is how many of your free users ever reached the thing you charge for. If you need the site side of the funnel too, the SaaS website conversion rate post covers visitor to signup, which is a different metric with different fixes.
A transcription app with one file per account
Picture an AI transcription app for people who record interviews: reporters, user researchers, podcast hosts. The free plan gives you 60 minutes of audio a month. Pro removes the limit and adds two things, speakers the app remembers and search across every file you have ever uploaded.
After signup the first screen is a dashed box that reads Drop an audio or video file, with MP3, M4A, WAV or MP4 up to 2 GB under it in grey. You drop in a 38 minute interview. A bar reads Transcribing, and about three minutes later the transcript appears. It is good. Paragraphs carry timestamps. The two voices are labelled Speaker 1 and Speaker 2. Top right sits an Export button with a small lock, and its tooltip says Word and SRT export are on Pro. The left sidebar shows All files (1), Search all files and Speakers.
Here is what the screen never says. If you rename Speaker 2 to the person's real name once, the app will find that voice in every later recording. Search all files is the reason a reporter with forty interviews would pay. Both features do nothing until there is a second file. The transcript page ends with the last paragraph of text and no next step.
Now the counts for one month of signups, in this example. 2,400 people make a free account. 2,160 upload a first file. 864 upload a second file within two weeks. 1,536, almost two in every three, never do. 50 people pay within a month, which is about two in every hundred. 48 of those 50 had uploaded at least two files.
Two plan table rewrites in one quarter
The founder opens Stripe every morning before standup. The internal dashboard has a tile labelled Free to paid, 30-day, and it has read about two in a hundred for four months. He knows the low end of the freemium range when he sees it, so he works on the part of the funnel that shows up in his tools: the upgrade.
In the first rewrite he cuts the Pro price by a fifth and adds a comparison table with fourteen rows of checkmarks. Three weeks later the tile has barely moved. In the second rewrite he cuts the table to five rows, adds a third plan called Studio for shared folders, and writes an FAQ under it about what happens to files when you downgrade. He posts the new page in a founder Slack and gets thirty replies, almost all of them about the price. Each rewrite took a designer and an engineer most of two weeks.
Nobody has opened the transcript page since launch to ask what a new user is meant to do after reading it. He knows the product too well to see that gap. To him, a second file is obvious, because he has uploaded hundreds. The person on the other side uploaded one interview to see if it worked, read it, and closed the tab.
Split the rate into the two numbers inside it
Free-to-paid is two numbers multiplied together. The first is the share of free users who reach real use, which in this app is a second file. The second is the share of those people who then pay. Work it through for the example. 864 of 2,400 reached a second file, a bit over one in three. 48 of those 864 paid, about one in eighteen. The other 1,536 produced two payers between them.
Look at where the pricing page sits in that sum. It can only touch the second number, and only for people who got far enough to want Pro. Both plan table rewrites were aimed at 864 people. The 1,536 who left after one file never saw the new table, and the old one was never their reason for leaving.
This is the part most conversion advice skips. It treats free-to-paid as a pricing and paywall metric, so it hands you tests for button copy, price anchors and trial length. Those tests are worth running once the first number is healthy. Run them before that and you are tuning the last step of a funnel that most people leave at the first one.
If you are not sure what your own second file is, start with how to define activation for an AI product. The short rule: the event has to be the user doing something with an output, and it should come before the paywall, not after it.
What the one-file accounts cost an AI company
In a normal SaaS product a free user who leaves is a lost acquisition cost. In an AI product that user also cost you compute. Every one of those 1,536 single interviews was transcribed by a model you pay for. AI product builders average around 52 percent gross margin, against 70 to 80 percent for traditional software, because inference sits in cost of goods. So the free plan that never turns into a habit is not only failing to earn. It is spending.
Then compare the two levers. Product-led SaaS benchmarks link a ten point gain in activation to roughly 15 to 25 percent more free-to-paid. In this example that means moving second-file users from about 36 to about 46 in every hundred signups, and going from 50 payers a month to somewhere around 58 to 62. That is a modest-looking number. It is still more than both plan table rewrites produced together, and it compounds, because the people who reach file two are the ones who stay.
There is a fundraising cost too. An investor looking at this company will ask for the retention curve, not the pricing page. A curve built from accounts with one file in them falls fast and stays low. That curve is what gets priced.
Five checks before you touch price
You can do all of these this week with the database you already have and an afternoon of watching sessions.
- Name your real-use event in one line. For the transcription app it is uploading a second file within two weeks. Write yours down with a count and a time limit.
- Split last month's signups into those who reached it and those who did not. Then count payers in each group. If almost every payer is in the first group, your rate is an activation problem.
- Read the last screen a first-time user sees, word by word. In the example it is the end of a transcript. Ask what it tells a stranger to do next. If the answer is nothing, that screen is your first fix.
- Put the paid value on the free screen. Here that means a line under Speaker 2 that reads Who is Speaker 2? Name them once and we will find them in every file, plus a button that opens the upload box again. Show the locked feature doing its job, not just a lock.
- Watch five recordings of new users from their first session. Note the moment each one stops. Do not change the price until you can say where those five people left and why.
Then wait one full monthly cohort. Compare the share who reach the event, and the payers inside that group, against the month before.
When the pricing page really is the problem
Sometimes it is. If a healthy share of free users reaches real use, inside the product-led activation range of 20 to 40 percent, and very few of them pay, the gap is after activation. Maybe the free plan gives away the whole job. Maybe the paid plan is priced in units nobody can turn into their own work, which the AI product pricing page post deals with. Maybe the upgrade prompt only appears on a settings screen nobody visits.
The decision rule is short. Low share reaching real use: fix the first run, leave the price alone. Healthy share reaching real use but few paying: now test the paywall, the plan limits and the page. Most founders at seed start with the second fix while sitting in the first case, because the pricing page is visible and the empty space after a first result is not.
Studio Maydit designs the first run, empty states and upgrade moments for AI products, which are the screens that decide whether free users ever reach the part you charge for. We rewrote Dualite's product copy screen by screen and rebuilt the moments where users stalled, and Dualite reached 100,000+ users in seven months, as the Dualite case study shows. Every fixed-scope project ends with a diagnosis of where the product is leaking, and you can see what that work covers on our product design page. If your free-to-paid has stayed flat through more than one pricing change, book a 30-minute call with Studio Maydit.
