Skip to main content

Case study

Our first case study is our own site

Ranksify is new, so instead of borrowing someone else's logo we are showing you the only account we can speak about without permission or exaggeration: ours. This page is the method — what we track about ranksify.ai, which engines answer, how the numbers are computed, and how you would run exactly the same thing on your own domain.

Disclosure

This is our own account, and we are not the neutral party

Treat everything on this page the way you would treat a vendor grading their own homework — which is why what follows is a method you can reproduce rather than a result you have to take on trust.

The subject

ranksify.ai — a new domain with no backlink history and no brand recognition to inherit. That makes it a hard case, not a flattering one: there is nothing here that a reader with an established site could dismiss as a head start.

The operator

Us. The same team that wrote the scoring code reads the dashboard, which is precisely why the scoring code refuses to round a small sample into a confident-looking number.

The claim

None, yet. We are publishing the instrument before the reading. When we do publish figures they will carry their sample sizes and intervals, the same as yours do.

What we track

The four things we watch about ourselves

This is the same setup any brand gets on day one; nothing here is an internal-only view.

  1. 01 · Prompts

    A tracked set of the questions a buyer would actually type — “how do I see if ChatGPT mentions my brand”, “AI visibility tracking tools”, and their unbranded cousins. Each one is re-asked on a schedule so the result is a series, not an anecdote.

  2. 02 · Answers

    Every answer is stored whole — the prose, the brands named, the sources cited and the position we appeared in. When a number on the dashboard looks wrong, the raw answer behind it is one click away, which is the only real defence against a metric drifting away from reality.

  3. 03 · Citations

    Which pages of ours get cited, which competitor and editorial pages get cited instead, and which of our pages have never been cited at all. The last group is the work queue.

  4. 04 · AI crawlers and the traffic they send

    A beacon on our own site records which AI crawlers fetch which pages, and which referrals arrive from an assistant afterwards. It stores no visitor IP address, so the country breakdown you would get from a conventional analytics tool is one thing we deliberately do not have.

Which engines

Five engines, asked directly

Answer engines disagree with each other constantly, so a single-engine reading is not a visibility measurement — it is one model's mood.

EngineHow we ask it
ChatGPTThrough the model API, not by scraping a UI.
ClaudeThrough the model API.
GeminiThrough the model API.
PerplexityThrough the model API, which returns its sources with the answer.
Google AI OverviewsCaptured from the search result, because there is no API for it.

The statistics

Why our own numbers are less flattering than they could be

Every rate on this product is a proportion measured from a finite number of samples. Treating that proportion as a fact is the single most common way an AI-visibility dashboard lies to its owner — including to us.

Wilson intervals, not point estimates

Every mention rate, citation rate and attribution rate is published with a 95% Wilson confidence interval. Wilson rather than the textbook normal approximation because it stays honest at small sample counts and at rates near 0% and 100% — exactly where a new site like ours lives.

Sample size travels with the number

A rate is never shown without the count it came from. Two of five is not 40%; it is two of five, and the interval around it is wide enough to say so. Most of the movement people celebrate in week one is inside that interval.

Benchmarks come from your own cohort

When we compare ourselves to a competitor, the comparison is against the competitors this brand tracks and has actually sampled. There is no cross-customer population in this product, so there is no “versus 10,000 brands” number for us to quote at you.

The practical consequence is that our own dashboard tells us to wait more often than it tells us we have won. We would rather ship that than a product that congratulates its owner on noise. The maths is in the documentation, not behind a sales call.

Reproduce it

Run the same method on your own domain

Four steps, in the order we did them. None of it requires our help.

  1. Step 1 · Write the prompts a buyer would type

    Start unbranded. “Best tool for X” tells you whether an engine has ever heard of you; “is [your brand] any good” only tells you what it says once it has. Both matter, in that order.

  2. Step 2 · Sample on a schedule, then leave it alone

    Answers vary run to run for reasons that have nothing to do with you. Collect for long enough that the interval narrows before you conclude anything. This is the step everyone skips.

  3. Step 3 · Read the citations, not just the score

    The score tells you where you stand; the cited sources tell you why. If the same three third-party pages are answering your category, your next piece of work is on those pages, not on your own homepage.

  4. Step 4 · Watch the crawlers, then the referrals

    A crawler fetching a page is the leading indicator; an assistant sending a visitor to it is the lagging one. The gap between the two is the honest measure of whether your content earned its place in the answer.

Start free trial

Prefer to check our reasoning first? The comparison with Searchable is sourced line by line.