AIO Case Study Guide: How to Measure Before and After AI Visibility Gains

A before-and-after case study is only as strong as the baseline and the log of changes behind it.

Every agency now claims to have made some brand visible in ChatGPT, Perplexity or Google AI Overviews, and almost none of them can show you the morning before it happened. That absence is the whole problem. So the first question to put to any case study, your own included, is a plain one: where is the dated record of the day before?

I have come to think of these documents rather as one thinks of a county's land records: dull to read, and yet the only thing standing between a rightful claim and a convenient fiction. When we grew HelpSquad's inbound leads, the figure that persuaded people was not a screenshot of a chatbot but a plain count, a handful of leads a month becoming a handful a day, set against a count taken before the work began. Our own guarantee is written against a client's target question set for the same reason. Questions fixed in advance cannot be quietly swapped for easier ones.

What follows is the method, step by step, beginning where every honest study must.

What is an AIO case study, and why do most of them fail to convince?

An AIO case study is a dated record showing that specific content changes moved a brand's presence inside AI answers, measured against a fixed baseline rather than against memory.

Before a single page is rewritten, settle four things in writing:

1. Name the questions you will measure, and freeze the list.
2. Name the engines you will sample, such as ChatGPT, Perplexity and Google AI Overviews.
3. Date the baseline and keep the raw export, not a summary of it.
4. Decide which business outcome will count as the result.

It is strange how much of what now passes for proof in this trade would not survive ten minutes with a sceptical finance director. The company I founded before this one, LiveHelpNow, sat on the INC 5000 at #84 (2015-2018), and growth of that order is judged by ledgers, with a before, an after and plain arithmetic between them; a screenshot of one chatbot reply, taken on a lucky afternoon, is the commercial equivalent of a postcard from a country the sender has never visited. Leadership knows this, even when it is too polite to say so.

The common assumption is that a brand which ranks well in Google will be cited by the AI engines as a matter of course. In 2025, Ethan, describing a study of thousands of questions on Lenny's Podcast, put ChatGPT's citation overlap with Google results at around 35% and Perplexity's at around 70%. Rank, therefore, is not a proxy. Each engine has to be measured on its own terms, and a case study that reports one engine while implying all of them has told a half-truth with the confidence of a whole one.

We have scored 11,000+ domains across 15 sectors and 28 categories, and the reason to score so widely is that a baseline means something only when it stands beside its peers. A number without neighbours is decoration.

The unit of measurement matters as much as the instrument. Our own guarantee is written against a client's target question set over 100 days, and I would urge the same discipline on anyone assembling evidence of AI visibility gains: the questions are chosen before the work begins, never after the results arrive. A question set selected in hindsight is not evidence. It is autobiography.

An analysis of 4 sources in this section, from a podcast study to our own scoring records, converges on one verdict: proof begins with a frame fixed in advance. That principle sits underneath all research-driven content built on evidence, and it applies with double force to claims about the engines themselves, which brings us, inevitably, to the baseline and to the morning on which it must be taken.

How do you set a baseline before you change anything?

A defensible baseline freezes the questions, names the engines and records each page's readiness before any edit; an audit, of which we have run 26,577, is simply that record taken seriously.

The difficulty, and it is a real one, is that the thing being measured will not hold still. An AI answer is less a fixed monument than a market square at different hours of the day, crowded at one visit and nearly empty at the next, so a baseline taken once and filed away describes a single hour and pretends it describes the town. For this reason the baseline must be built to survive repetition, which means it is a procedure and not a photograph:

• Question set: the exact wording of every prompt, saved as text, with the date it was fixed.
• Engines and conditions: which engines were asked, whether the session was logged out, and how many times each prompt was run.
• Presence record: whether the brand was named, whether it was linked, and which rival domains appeared beside it.
• Page readiness: a Rank for each page you intend to change, taken on the same day.
• Business state: inbound leads, booked calls and referral sessions for the same period.

That last item is the one most often forgotten. Consider the plainest of our results, a client who went from a quiet site to 10 booked calls in 3 weeks: the word quiet is doing as much work in that sentence as the number ten, and had nobody written down how quiet the site was, the ten calls would have floated free of any comparison, an effect with no recorded cause and no recorded starting point. A result is only as strong as the before it is measured against.

There is a craft to this which looks humdrum and is not. Our team includes a full-stack engineer and content infrastructure architect with 20 years of building enterprise systems, and the habits of that discipline (versioned inputs, timestamped outputs, nothing overwritten) are exactly the habits a baseline needs. Engineers learned long ago that a log nobody can replay is a rumour.

The same rigour applies to what you claim a page can answer. In our repo-based knowledge base pilot, the study step finished a two-topic study in about two minutes, each proposed topic cited one to four specific files and line ranges that answer it, and the study dropped a topic when the controls it would describe were not yet wired up in the code. A baseline question you cannot trace to a real source is one you should strike. Readers who want the vocabulary behind these checks will find it set out in our AEO FAQ on AI search optimization.

And yet a baseline, however careful, is only half an argument. It tells you where you stood; it cannot tell you which numbers, gathered afterward, deserve to be believed, and that question turns on the instruments themselves.

Which metrics should a before-and-after AI visibility report use?

The result worth promising is being named in AI answers in 100 days, proven with platform impressions, repeated prompt checks and referral data rather than a single screenshot.

The instruments have improved this year, and the improvement is not trivial. CMSWire reported that Google launched dedicated Search Console reports on June 3, 2026 isolating site impressions inside AI Overviews, AI Mode and Discover's generative features, with five data dimensions (impressions, pages, countries, devices and dates), though rollout began with a subset of websites. The same report carried a warning that every case-study author should pin above the desk: of those five dimensions, impressions are the only measure, so nothing in the new reports says whether anyone clicked. Before trusting any chart, ask which of the four sources below produced it. Then put the same question to every vendor deck, as a surveyor puts it to a handsome facade: which room of the house does this window actually show?

No single source carries the whole argument, so I would assemble the report from four, each doing the one thing it can honestly do:

The bottom row is where a case study earns its keep. When we grew HelpSquad's inbound leads from about 5 a month to 5 a day, and the client came to serve 200 active clients, the impressions were the scaffolding and the leads were the building. Leadership will ask about the building. Investors will ask about nothing else.

Two cautions follow. First, keep every other intervention in the log, because a paid placement running in the same window (for example, sponsored listings that AI engines cite) will lift the same numbers and must be reported as a separate cause. Second, read the four sources together, never one at a time. Michael and Alex bring a combined 40+ years of SEO and content infrastructure experience to AEO, and if that long acquaintance with search teaches anything, it is that a metric admired in isolation will, sooner or later, be gamed by the very team that reports it.

Which raises the harder matter of what, precisely, was changed, and on which day.

What will a credible AIO case study look like a year from now?

It will lean on platform-reported data and logged changes rather than screenshots, because the engines are shifting faster than any single prompt can capture.

The ground is already moving. CMSWire, reporting on the Semrush AI Visibility Index, noted that brand mentions in Google AI Mode dropped 4% from August to October 2025 while source diversity rose 13%. More sources, fewer mentions each. A brand that held its place last quarter can lose it without changing a word, which is precisely why the case study must record what the brand changed and when. I expect Google's AI reports, still limited to a subset of sites at launch, to become the common starting point; and yet a platform's number is only a witness, not a verdict.

The verdict comes from the log. In our first repo-based pilot, a re-index after a day of commits re-read only the 14 files that changed. That is the standard I would hold every case study to. Know exactly what moved, and the arithmetic of before and after becomes something a board can trust.