The web has no memory. You are the polling loop.
NotifyMe watches one page against one sentence, remembers every observation, and speaks only when something worth acting on changes. This is the commercial case for building it past the hackathon.
SOURCE ISSUE, is the whole thesis.
The proposition
Every page you care about is a snapshot. The flight board, the rental listing, the resale ad, the job posting. None of them remember what they said an hour ago, and none of them tell you when it changed in a way that matters to you.
So people build a habit instead of a system. They refresh. They set a reminder to check. They open the same tab eleven times a day and still miss the moment, because the moment happened at 3am.
The tools that exist ask you to describe the change in their language: pick a CSS selector, set a threshold, choose a diff mode. NotifyMe asks for a sentence. Give it what you are waiting for, the page, and how often to look. It handles the rest, and it is honest with you about the times it could not look at all.
A failed extraction keeps the last good observation as visibly stale history. It never invents an absence.
The watch is one sentence: an NFT Enterprise Sales Executive role, remote, full time. The listing existed, then it was taken down, and NotifyMe said so out loud rather than letting the page go quiet.
Two details on this screen carry the whole argument. The status reads No match, not an error, because the page was read successfully and the role genuinely is not there any more. And the panel underneath is headed Real observations, because every check that produced this conclusion is stored and can be replayed.
Captured during the build. Web Push firing on a genuine disappearance event, not a rehearsed one.
Who pays for this
Three segments, ordered by how acute the pain is and how quickly they would pay.
Scarce-market buyers
Dubai renters, resale hunters, ticket buyers. Inventory is thin, moves fast, and the cost of missing it is measured in thousands of dirhams or a year in the wrong flat.
Small operators
Independent sellers and agencies watching competitor pricing, stock levels, and listing changes without an analyst or a scraping stack.
Job seekers
People watching a handful of specific company career pages, where being early is most of the advantage and job boards are always a day behind.
Five watches. One engine.
Four of the five live watches, captured on the day. A flight, a resale listing, a rental portal and an events site. Same prompt box, same pipeline, not one line of category-specific code between them.
Three of these four returned nothing at the moment of capture. That is the honest answer, and it is a different answer from "we could not look."
Why frequency is the price, not a feature gate
Every check costs us one Context.dev extraction. The gap between the cheapest and the most expensive tier is not a product decision, it is ninety-six times the work.
Every 15 minutes costs 96× what daily costs. Tiers track that curve instead of inventing limits.
The tier survives below roughly AED 0.005 per check and holds a healthy margin below AED 0.002. We do not yet know the real number, which is why it is the first thing we instrument.
The model
The market*
Built bottom-up rather than top-down, and Dubai first. We would rather be able to defend every step of the funnel than quote a continental number nobody can trace.
90,000 reachable*
× 3% free-to-paid conversion*
= 2,700 subscribers
× AED 29 × 12 months
= AED 939,600 ARR*
Under a thousand paying users in one city, in one year, on the cheapest paid tier. Pro and Team sit on top of that, and every additional category widens the top of the funnel without changing a line of the engine.
Conversion is the number we least want to guess at, which is why the first ninety days are spent measuring it rather than modelling it.
Who else is in this space
Why that third one is the whole argument
The first two columns are convenience, and convenience gets copied. Anyone can bolt a language model onto a diff tool and let you type a sentence. Anyone can widen a scraper to more categories.
What does not get copied is the discipline to admit failure. Visualping, Distill and Browse AI all report a page as unchanged when they could not read it, because reporting a technical failure to a consumer looks like a broken product. That decision is rational for them and catastrophic for you: an airline page that starts blocking requests goes quiet, and quiet reads exactly like "your flight is fine." Google Alerts has the same shape of problem from the other direction, watching the whole web for mentions rather than one page for a state.
NotifyMe treats "we could not look" as a first-class answer with its own status, its own persisted error code, and its own notification policy. It costs us a support conversation. It saves the user the thing they were actually watching for. Every other capability on this page is a feature; that one is a promise, and promises are what people stay subscribed to.
Our own assessment of the four capabilities that matter, not a vendor scorecard.
| Describe the outcome, not the change |
Works in any category |
Says when it could not look |
No setup beyond a sentence |
|---|
Only one row carries a yes in every column. That intersection is the product.
What could kill it
Extraction cost outruns consumer pricing. The whole model rests on a per-check price we do not control. Mitigation is frequency tiering, dormancy pausing, and change detection before extraction.
The sites people care about most block us. Airlines and portals are the most aggressive about automated access. We surface this as SOURCE ISSUE rather than hiding it, but a watch that permanently fails is a churned user.
Alert fatigue. If the meaningful-change threshold is wrong in either direction we are either noise or useless. This is the hardest product judgement in the system.
A platform absorbs it. Any large browser or assistant could ship a version of this. Our defence is the honesty model and per-category tuning, neither of which a generic feature tends to bother with.
The next ninety days
- Instrument real per-check extraction cost and replace every modelled figure on this page with a measured one.
- Multi-source watches, so one intent can cover three portals instead of three watches.
- Shared watches, which is the natural referral loop for flatmates, families, and job-hunting friends.
- A user-set alert budget: tell us how often you are willing to be interrupted, and we hold to it.
- Fifty paying users in Dubai before adding a single new category.
The figures marked with an asterisk are illustrative placeholders. They show the shape of the model and the order of magnitude we are working towards. They are not researched, sourced or observed numbers, and they should not be quoted as such.
- The market — total addressable, serviceable and year-one reachable counts
- What one Dubai year looks like — conversion rate and the resulting ARR
Three real inputs replace all of them: the Context.dev per-extraction price, a cited count of active property and resale searchers in the UAE, and an observed free-to-paid conversion rate. Instrumenting the first of those is task one of the next ninety days.
Everything else on this page is real: the checks-per-month arithmetic, the margin curve, the four screenshots and what they returned at capture, and our own assessment of the competitors.