The upgrade nudge that writes itself: convert free users before they hit the wall
By the time a user hits your paywall, they are blocked, annoyed, and halfway to a competitor’s signup page. The best moment to make the pitch was three days earlier - when they were winning. Here is how to catch it, automatically, for every user at once.
Last updated: 2026-06-05
The paywall is a hostage negotiation
Think about the emotional state of a user who just hit a hard limit. They were mid-task. Something they needed is now behind a wall, and the wall is asking for money. Some pay; many quietly leave. Every conversion study lands on the same conclusion: pitching at the moment of friction converts worse than pitching at the moment of success, because at the friction moment you are negotiating with someone you just blocked. The user at 80% of quota is a different person. They are getting value right now - the meter proves it - and nothing is interrupted. That is the user you want to talk to. The problem is that talking to them well requires knowing what they are doing, which limit they are approaching, and what users like them care about - per user, at scale. That used to mean a segmentation project. Now it is one compose type.
Find the 80% moment with the meter you already have
If you meter with Vevee, you already have the trigger. usage(userId) returns every counter for the user’s current period - quota, count, remaining - so detecting "this user just crossed 80% of their generations" is arithmetic on data you are already writing. Run the check right after track() on your server (you have the fresh counters in hand), and when a user crosses the threshold for the first time in a period, that is your moment. No new infrastructure, no analytics export, no nightly batch job scoring "PQL likelihood" in a spreadsheet.
// After your normal track() call - the moment detector
const { counters } = await vevee.usage(userId);
const gens = counters.find(c => c.event === 'image_generation');
const crossed80 =
gens && gens.quota > 0 && gens.count / gens.quota >= 0.8;
if (crossed80 && !(await nudgeAlreadyShownThisPeriod(userId))) {
await queueUpgradeNudge(userId); // generate + show on next page load
}One compose type writes every user’s pitch
A compose type is configured once in the Vevee dashboard: an intent prompt, a set of data sources, and a structured output schema. For the upgrade nudge, the intent is something like "This user is approaching their plan limit while actively using the product. Write a short, specific, encouraging upgrade nudge that references what they have actually been doing and frames the paid plan as removing a ceiling on something going well. No false urgency." Then you enable the sources that ground it in reality. Every source resolves against the userId you pass, so the same type produces a different, personally relevant message for each user - with zero persona branching in your code.
- user_usage - which limit they are approaching and the real numbers ("38 of 50 images")
- user_events - what they were doing this week, so the pitch references their actual workflow
- conversion_signals - what tends to precede an upgrade for users of this app
- cohort_compare - where this user sits versus similar users ("top 10% of free-tier creators")
The call: typed output, one fallback, no exceptions on the happy path
On the server, with your secret key, you name the type and pass the user. The generic parameter types the structured output - the schema you defined in the dashboard is the contract, and your UI renders whatever comes back. Two non-success cases and both are cheap to handle: the user may have opted out of AI personalization (a status, not an error - render your static banner), and the workspace AI budget may be exhausted (a VeveeError with code ai_budget_exceeded - same fallback). The static banner you have today becomes the floor, never the ceiling.
import { createClient, VeveeError } from '@vevee/sdk';
const vevee = createClient({ apiKey: process.env.VEVEE_SECRET_KEY! });
export async function upgradeNudge(userId: string) {
try {
const res = await vevee.compose<{
headline: string;
body: string;
ctaLabel: string;
}>('upgrade-nudge', userId);
if (res.status === 'opted_out') return STATIC_NUDGE; // respect it, silently
return res.output;
} catch (e) {
if (e instanceof VeveeError) return STATIC_NUDGE; // budget spent, type renamed...
throw e;
}
}What it actually says (and why specificity converts)
Generic nudges get banner blindness because they read like ads. A grounded nudge reads like the product talking: "You’ve generated 38 of your 50 images this month - your best month yet. Pro removes the cap, so next week’s client work doesn’t wait for the 1st." The numbers are real and the user can verify them in their own usage bar, which is exactly what makes the sentence persuasive. A student grinding through exam season, a freelancer with client deadlines, and a hobbyist who just discovered your new model each get a different sentence from the same compose type, because their context blocks differ. You improve all of them at once by editing the prompt in the dashboard - watch the copy change in production without a deploy.
- Freelancer at 80% of images: "38 of 50 used - Pro means client work never waits for a reset."
- Student at 80% of tokens: "You’ve summarized 12 lectures this week. Pro covers the whole semester for the price of a coffee."
- Power user near a premium-model cap: "You clearly prefer the 4k model. Pro makes it your default instead of your treat."
Cache it, cap it, measure it
Three disciplines keep this honest. Cache: generate once per threshold-crossing and store the result on the user record - the nudge does not change between page loads, so neither should the model call. Cap: keep maxOutputTokens low in the type config; this is a headline and two sentences, not an essay, and a tight output schema enforces that structurally. Measure: every compose call returns usage.costMicroUsd, and you should capture an upgrade_nudge_shown analytics event when you render it. Put nudge-shown and checkout_completed in the same funnel and you get the only number that matters: incremental upgrades per dollar of generation cost. At typical costs, a single extra monthly subscription pays for thousands of nudges - but now you can verify that claim in your own dashboard instead of taking it from a blog post.
// Render-side: make the nudge measurable like any other funnel step
await vevee.analytics.capture({
distinctId: userId,
event: 'upgrade_nudge_shown',
properties: { variant: 'composed', thresholdPct: 80 },
});Why this beats rewriting your paywall
The paywall still matters - keep it, personalize it too if you like. But the nudge wins on selection, timing, and tone. Selection: it only fires for users who are demonstrably getting value (you cannot reach 80% of a quota by accident). Timing: it arrives while the user is succeeding, days before frustration. Tone: it is the product noticing your momentum, not the product holding your work hostage. Conversion work usually means rewriting copy and praying; this is the version where the copy writes itself from the user’s own meter, every fallback path is two lines, and the whole loop - trigger, generation, render, funnel - is observable end to end. The wall converts the desperate. The nudge converts the happy. There are more happy users.
More from the blog
The race condition in "if (usage < limit)" that is costing your AI app money
A user at nine of ten images opens six tabs and clicks generate in all of them. Every tab passes your limit check, every tab gets an image, and you pay for all six. The bug is one read-then-write - and your unit tests will never catch it.
engineering · 9 minHow to build a credits system for your AI app (ledger design, rollover, refunds)
A user emails: "I had 40 credits this morning and now I have 12, and I only made one image." If your balance is a column you mutated, you cannot answer that email. Here is the ledger design that can.
thinking · 7 minStripe metered billing for AI apps: what it solves and what it does not
You are sketching the pricing page for your AI app and every plan ends in "per generation" - because your costs arrive per generation. Stripe meters the charging half beautifully. The question is what happens at request time.
thinking · 8 minHelicone is in maintenance mode. An honest map of where to go
Somewhere in your codebase a base URL points at Helicone, and it has been quietly doing four jobs at once. The team got acquired - good for them - but now you have to replace each job separately, and the one everyone forgets is the one your users will exploit.
engineering · 8 minHow to stop free-tier abuse without killing signups
The usage graph spikes at 3:41am: four thousand free generations in an hour, all from disposable emails. Your first instinct is a credit-card wall. That instinct will cost you more than the abuse does.
thinking · 7 minTokens, credits, or requests: choosing the unit you meter (and price)
The pricing doc has three columns: $9, $19, $49. The prices took twenty minutes. The row above them - what a user actually gets for the money - has been blank for a week. That blank row is the real decision.
engineering · 3 minI rewrote my cancel flow with one LLM call. It argues better than I do.
The cancel flow is where SaaS revenue goes to die politely. "Are you sure? You'll lose access to Premium features" has never changed a single mind - but reminding a user of their own 312 generations this month is an argument.
engineering · 3 minThe trial-ending email everyone sends is the same email. Here's the one that converts
"Your trial ends in 3 days! Upgrade now to keep access." You've received a hundred of these. You've deleted a hundred of these. The email fails because it's about your product, not about the user's trial.
engineering · 3 minStop making users do math on your pricing page: recommend their plan from their own usage
Every pricing page asks the user to solve an estimation problem: "How many credits will I need per month?" Nobody knows. But for any user who has actually used your product, their usage history IS the answer - here is how to put it on the pricing page.
engineering · 3 minWin-back emails fail because they're written for "users." Write them for the one user instead.
Every dormant-user campaign in history: "We miss you! Here's what's new." Open rate: pity clicks. The email fails because it's about your changelog - and the user left because of something in their experience.
engineering · 3 minSpotify Wrapped is a growth loop, not a year-end gimmick. Ship one for your AI app in an afternoon.
Spotify Wrapped works because people love seeing their own behavior reflected back as a story. Every AI app with usage data can run that loop monthly - and almost none do, because turning usage rows into narrative used to be a content problem. It is now a schema problem.
engineering · 3 min"AI-personalized copy converts better" - prove it or delete it. Here is the 40-line A/B harness
Half of the "we added AI personalization and conversions went up 40%" posts have no control group. The other half measured clicks, not revenue. If LLM-generated copy is going in front of your paywall, you owe yourself a real experiment - and the harness is tiny.
thinking · 3 minAI personalization without the creepy part: opt-out as a first-class return value
Users have learned that "personalized for you" means "we mined everything you ever did" - and the backlash is rational. When I added LLM-generated personalized copy to my app, the part I sweated was not the generation. It was making declining it a real, respected choice.
engineering · 3 minI stopped walking into demo calls blind: every lead now comes with a usage brief written 10 seconds before the call
Founder-led sales has one structural weakness: you have no time to prep. If the lead has touched your product, their usage history is the best discovery call you'll never have to run - here's how I turn it into a one-page brief, automatically, before every call.
thinking · 4 minThe "founder email" converts like crazy and scales like garbage. Here is the middle path.
A personal email from the founder converts trials at a rate no automated sequence touches - and at 50 signups a week it stops scaling. The middle path: drafts generated from each user's real usage, that you read, edit, and send yourself.
engineering · 3 minChurn doesn't announce itself. My Monday Slack digest does it instead.
Every founder finds out about churn the same way: the cancellation email. By then the user has been gone for weeks - the decision happened earlier, quietly, in their usage. So I made the metering tables write me a memo.
engineering · 3 minHalf of every support ticket is asking what the user already did. Attach the answer instead.
The ticket says "it's not working" - and the next twenty minutes go to figuring out who this user is, what plan they're on, and whether "it" is a bug, a quota, or a misunderstanding. The actual fix usually takes two. All of that context lives in your usage data; here is how to attach it to every ticket automatically.
thinking · 4 minYour users are telling you your roadmap, in writing, every day. It's in your prompt logs
Founders pay for interviews and beg for survey responses to learn what people are actually trying to do. Meanwhile, your users type their intent into your product, in their own words, hundreds of times a day - and nobody performs for a prompt box.
engineering · 5 minHow to reset usage limits when a subscription renews
A weekly plan, used twice. First week: ten generations. Second week: zero, because the counter never reset. This is the cron-job mistake - and the fix is one field on one call.
engineering · 6 minHow to manage subscription renewals: aligning Vevee with Stripe
A user signs up on Jan 15. Stripe charges them on the 15th of every month. Your metering layer resets on the 1st. Two clocks. One angry support ticket per cycle.
engineering · 5 minStop hardcoding your pricing page - render it from your metering layer
Every B2B SaaS I have shipped repeats the same mistake: plans live in two places - the dashboard that enforces them, and a const PLANS = [...] on the marketing site. They drift within a quarter.
thinking · 4 minMeter AI by user, not by account - your margin depends on it
A few users will cost you 100x what your median user costs. If you only meter at the account level, you will not see them coming until your gross margin is gone.
engineering · 5 minreserve / commit / release: the only correct way to enforce AI quotas
Every team I have seen build per-user AI metering has shipped a version of canUse → call OpenAI → track. It looks correct in single-threaded tests. It is broken in production.
thinking · 4 minWhy Stripe Billing is not enough for AI products
Stripe is excellent at one thing: turning usage into invoices. AI products need three other things, and Stripe does not do any of them.
engineering · 6 minDynamic onboarding: a different first step for every user
A teacher and a student sign up the same minute. The teacher wants to build a quiz; the student wants to summarize a lecture. Your onboarding shows them both the same five-step checklist. One of them bounces.
thinking · 6 minPaywall copy that rewrites itself for every user
Your paywall says "Upgrade to Pro for unlimited generations." A teacher reads it and shrugs. A student on a budget reads it and closes the tab. The same words, two lost conversions - because the words were written for nobody in particular.
engineering · 5 minAdd usage limits to your AI app in 10 minutes (no backend required)
You shipped an AI feature on Friday. By Monday one user had burned $212 of OpenAI credit on your free tier. The fix is not a TODO comment that says "add rate limiting" - it is two method calls.
engineering · 5 minMeter LLM tokens, not requests - your flat per-request limit is lying
Two users, both at 100 requests. One sent tweets, the other sent novels. Your cost for them differed by 400x. Your limit treated them identically - and your margins noticed.
engineering · 5 minOne event, two limits: gate your premium model without forking your code
The premium model launch was going great until you looked at the bill: free users had figured out the good model and were living on it. You need a sub-limit. You do not need a second code path.
engineering · 5 minTest mode: break your pricing in the sandbox, not on your customers
You changed the free tier from 10 to 25 generations and somehow locked out every Pro user for an hour. Nobody tested it, because testing it meant tracking fake events into production analytics. There is a mode for this.
engineering · 5 minThe support ticket that solves itself: log the prompt behind every AI event
A user says your AI feature "broke" on Tuesday. You have a charge for the call, a timestamp, and no idea what they asked or what the model said. The evidence existed for exactly one request - the one you didn’t log.
engineering · 5 minYour funnel has one broken step. Find it without writing a single SQL query.
A hundred people saw your paywall this week. Three upgraded. Is that a copy problem, a price problem, or did ninety of them never generate anything worth paying for? You cannot fix what you cannot locate.
engineering · 5 minOne user, three ghosts: fix your funnel with identify()
Your funnel says signup conversion is 4%. It is actually 11%. The missing users didn’t bounce - they came back on another tab and got counted as someone new. Every number downstream of that split is wrong.
thinking · 5 minThe usage bar that sells the upgrade (build it with a public key in an afternoon)
An invisible limit feels like a trap. A visible one feels like a fuel gauge. Same quota, same plan, same user - and a measurably different reaction when the wall finally arrives.
engineering · 6 minHow to cancel a subscription without burning the bridge (or your data)
The user cancels. Do they drop to a free tier, or lose access when the paid month runs out? Those are different products, different SDK calls, and different mistakes when you get them wrong.