No certifications.
A published method instead.

Citino has no SOC 2 report, no ISO certificate and no compliance badges, and this page is not going to imply otherwise with a row of grey logos. We are new, and those take time and money we have not spent yet. What a badge buys is a stranger's assurance that we do what we say. In place of it, here is the thing itself: every decision that turns an engine's answer into a number on your chart, written out with the mechanism, so you can check the arithmetic rather than trust the auditor.

This is the second most-read page on the site, and it should be. Every other page shows you a number. This one tells you how wide the number's error is, what it excludes, when it is allowed to alert you, and what happens to your history when we improve the parser. If any of it turns out to be wrong, the corrections policy is further down and it is binding on the public benchmark pages as well as on your dashboard.

Read it adversarially. The whole point of publishing it is that you can.

Treat all 1,785 answers as 1,785 independent observations, resample them, take the middle 95%. It produces a tight, confident-looking band: on Tallyfold's 10.3%, roughly ±1.5. It is wrong because those answers are not independent. Thirty answers to "best warehouse management software for a 3PL" are near-copies of each other — if an engine names Crateline first today it very probably names Crateline first tomorrow. Repeating a question does not tell you more about the category; it tells you more about that one question. Counting each repeat as fresh evidence inflates the denominator without adding information, and because the interval narrows with the square root of the effective sample, the band comes out about half as wide as it really is.

Draw 54 prompts with replacement from the 54, and each time a prompt is drawn, recompute the metric using all of that prompt's answers as one block. Repeat many thousands of times, take the 2.5th and 97.5th percentiles of the resulting distribution. The prompt is the independent unit, because a new prompt is a new question and a new answer to an old prompt mostly is not. On Tallyfold's 10.3% this gives ±3.0, and that is the number the product prints.

How the interval is computed is the single most consequential decision in the product.

Every share Citino prints is a proportion drawn from a finite sample, so it carries a 95% interval. Most tools get this wrong in the direction that flatters them.

The difference is not academic. A band that is half as wide as it should be turns every ordinary week's wobble into a movement that clears its interval — which is exactly how a tool ends up emailing you about noise, and exactly why Tallyfold's +1.7 is grey here and would be green almost anywhere else.

Illustrative. The ±1.5 is what the wrong arithmetic prints on the same data, not a Citino number.

Tallyfold gained 1.7 points over thirty days on a band of ±3.0. Citino labels that within noise and does three specific things with it.

It prints it grey, with no arrow and no colour, because colour is a claim.

It does not alert on it. Alerts fire only on a delta that clears its own interval, so an empty alert inbox means the week was quiet rather than that the job is done.

And it never sorts as a movement. The biggest-movers list is computed only over deltas that cleared their bands, so a within-noise row cannot reach the top of it by being the largest of the small numbers. If nothing cleared this week, the list is empty and says so.

Engines change. A model is updated, a retrieval policy shifts, a system prompt is rewritten, and the whole category moves on a Tuesday for reasons that have nothing to do with anything you published. Two rules handle it.

Drift is labelled, and never dropped. When a category-wide shift is detected, the date is marked on every chart in that category as a labelled drift event with the engine named. We do not smooth it, we do not restate history to make the line continuous, and we do not quietly re-baseline so that last month looks different than it did last month.

Your change is reported net of the category baseline. Rewriting the 3PL billing page took Tallyfold from 21.1% to 29.5% — a raw +8.3. Over the same window the whole category moved +0.5, so the reported effect is +7.8. Without the baseline you cannot tell a good page from a good week.

A prompt cannot be edited after it starts collecting — only retired and replaced. New wording is a new prompt with a new id, and it enters the chart as a hard break rather than a continuation, because the two wordings are different questions with different answer sets. The prompt set carries a version, and every archived answer records the version it was collected under.

This costs us something. When a prompt turns out to be badly worded, we cannot fix it and keep the history; we retire it, write the better one, and wait for the new series to be worth reading. Editing a prompt in place is how a chart silently becomes fiction.

The engines' own official APIs, on paid accounts, at a fixed daily schedule. Every answer we report was returned to a request we made, under terms that permit it, and we can show you the response.

No scraping of consumer interfaces. No headless browsers driving a logged-in chat window. No third-party search proxies, no resold "AI rank" feeds, no data bought from someone who does not say how they got it.

Real coverage. Copilot and Google AI Overviews are not sampled at all, because neither offers an official answer API — and both matter to buyers. Where they appear anywhere in Citino it is as a labelled referral source in your GA4 attribution, never as a measured answer. Adding a new engine takes as long as that engine takes to ship an API, so we are sometimes late. And the bill per answer is higher than scraping, which is part of why custom prompts are the thing the plans meter.

A scraped number and a sampled number look identical on a chart. If we mixed them you would have no way to tell which of your lines was built on a source that could break, change silently, or be turned off by the engine at any moment — and neither would we, six months later, when the archive is the only record of what happened. A measurement is only as durable as its least defensible input.

Personalised answers. We sample clean sessions on a fixed schedule. Your buyer has chat history, an account, a location, a time of day and a way of phrasing things, and every one of those moves an answer. What Citino measures is what the engines say to a neutral asker of a fixed question — a stable, comparable population, and not a transcript of what any specific person saw.

Logged-in and enterprise surfaces. Not sampled. Neither are Copilot and Google AI Overviews, for the API reason above.

Twenty-nine per cent of your AI-referred sessions. Of 1,240 AI-referred sessions in thirty days, 71% can be matched to a prompt we sampled. The remaining 29% cannot, and the product reports that 29% as its own line rather than distributing it across the matched prompts. Distributing it would make every prompt's value look 40% larger and the report look complete.

Illustrative. Figures from the sample category.