Guide · Evidence

How to read the numbers in an AI agent case study.

Matched windows, paid traffic removed, one-day snapshots. The checks we run on any results page, including our own.

In short
  • Find the comparison first. The same months a year apart beat month-on-month.
  • Check what was removed, especially paid media, and whether there is a window before any spend.
  • Check the unit. Audit scores, traffic, bookings and revenue are different claims.
  • Check the base behind each percentage, and who measured it.
  • Treat a one-day ranking or AI-answer check as a snapshot, and ask for the second measurement.
  • Read the limits line before the headline.

A percentage with no comparison is decoration. Before you trust a case study, check what it compared, what it left out and what it admits it cannot show. These are the checks we run on our own case files, with worked examples from them.

1. Find the comparison

The strongest comparison is the same period a year apart, because seasonality drops out. The Arros QD three-month report compares each of July, August and September 2026 with the same month of 2025: 10,422 sessions became 27,911, up 167.8%.

Month-on-month comparisons are weaker. Rocket Room's monthly sessions rose 146.9% from June to September 2026, from 1,278 to 3,155. The case file states its own limit: there is no year-over-year baseline, so the season is still inside that number.

2. Ask what was removed

Paid media adds traffic on its own. The Arros QD report takes Paid Search, Paid Social, Display and Cross-network out of each month's channel table, and sessions still rose 128.5% in August and 103.5% in September. A result that survives that subtraction is not a spending result.

Where paid media was not removed, a good case study says so. Rocket Room's file notes that Paid Social and Paid Search also ran over the period.

3. Look for a window before the spend

Arros QD's site launched on 26 June 2026, and Paid Search began on 6 July. The nine days after launch and before paid media, set against the nine days before launch, form a window with no ad spend in it. Sessions rose from 905 to 2,751, up 204%.

4. Check the unit

A case study can report five different things, and they are not interchangeable. Each step down the list is a stronger claim and harder to measure:

UnitWhat it showsExample from our files
Audit scoreWhat a tool measured about the siteWhite Dream Villas: core indexing 41 to 100
VisibilityRankings, AI mentions and citationsArros QD: cited directly by Perplexity
TrafficSessions, users, engagementRocket Room: sessions up 146.9%
ActionsCalls answered, bookings, enquiriesEnvy Smile: 129 bookings from 296 calls
MoneyRevenue or cost savedNot reported in these files

Each file says what it does not reach. White Dream Villas reports audit scores, not traffic or bookings. Rocket Room tracks no booking events, so conversions are not reported. Envy Smile reports bookings, not revenue per booking.

5. Check the base

A percentage needs its base. Envy Smile's 44% conversion rate is 129 bookings from 296 calls answered over six months. 146 of those calls came outside surgery hours, and monthly conversion rose from 20% in December to 56% in May. With the base you can see the size of the sample and the trend inside it.

6. Check who measured it

Find the source of each number:

All three are useful. They are not the same strength, and a good case study tells you which one you are reading.

7. Treat a snapshot as a snapshot

An AI visibility check usually asks several assistants, such as ChatGPT, Gemini, Claude and Perplexity, a set of questions a customer might ask. It records whether the business was the top pick, shortlisted, mentioned in passing or absent, and whether the answer linked to its website. That is useful, with three caveats:

Rankings have the same problem. A third-place result can be the seventh thing a searcher sees once ads and Google's own boxes are counted, and the page that ranks may be the homepage rather than the page you want people to land on. Check which page ranks, and for which search.

Ask for the second measurement, with the same questions, the same assistants and the same searches, after the changes.

8. Look at how the data was cleaned

Real analytics are noisy. The Arros QD report states that GA4 recorded unusually few sessions from 24 August to 13 September, and that those days are included as recorded, which makes August and September conservative. It flags one suspicious referral that sent 480 sessions on 24 July; without it, July is still up 192.0% on July 2025.

It also filters out sessions recorded on the booking provider's domain after the booking widget was embedded: 4,854 of them in September alone. If a page shows no sign of cleaning, ask what was left in.

9. Read the limits line first

A page that names its limits tells you where the number stops. Every ZeroLayer case file carries one: what the agent did, what changed, and what the data cannot yet show. Read it before the headline figure, and be wary of a results page that has none.

The checklist

QuestionA good answer
What was it compared against?The same period a year earlier, or a matched window
What was removed?Paid channels and known spam, stated on the page
What is the unit?Named: audit score, visibility, traffic, actions or money
What is the base?Counts shown alongside percentages
Who measured it?The client's analytics or logs, with the source named
Is it a snapshot?A second measurement on the same questions
What are the limits?Written on the page

All six of our case files, with their limits, are on the case studies page.

Every ZeroLayer case file carries its limits. What the agent did, what changed, and what the data cannot yet show.
See the evidence →
FAQ

Common questions

What makes a case study trustworthy?

A stated comparison period, a clear account of what was removed or excluded, a named unit and base, the source of each figure, and a limits line.

Why compare against the same month last year?

It holds the season constant, so a change in sessions reflects the site and not the calendar.

Is a one-day AI visibility check reliable?

It shows where a business stood on one day, for the questions asked. It cannot show a trend until it is repeated with the same questions and assistants.

Where do the figures in ZeroLayer case studies come from?

From the client's own analytics, call logs or estimates, and each page says which. Each case file carries a limits line.

Planning an agent build?