In partnership with
2026 State of AEO Report
A year ago, most marketers weren't thinking about AI search. Now it's one of the fastest moving channels in the industry and nobody has a playbook yet.
So we built one. We surveyed hundreds of marketers to find out how they're approaching answer engine optimization, where they're investing, what's actually working, and what isn't.
The result is the 2026 State of AEO Report. Real data. Real strategies. A clear picture of where AI search is headed and how to get ahead of it.
Every growth team has the same recurring moment.
A number moved. Conversion dropped 12 percent last month. Or signups are up, but revenue is flat. Or a channel that was working stopped working. Something in the data is wrong, and now someone has to figure out why.
What usually happens next is not analysis. It is a guess. Someone in the meeting has a theory. Someone else has a different theory. The theories get debated based on whoever is most senior or most confident, a decision gets made, and the team acts on a hypothesis that was never actually tested against the data. Sometimes the guess is right. Often it is not, and the real problem keeps running for another two months until it becomes too big to ignore.
The reason teams guess instead of analysing is not laziness. It is that real diagnostic analysis, the kind that pulls the data apart from six angles and finds the actual cause, that takes hours of focused work that a busy team does not have. So they shortcut it with intuition.
Claude removes that shortcut because it collapses the hours into minutes. This edition is about using Claude as a diagnostic partner, the data detective that takes a number that moved and tells you, with evidence, why.
Why this works, and the one rule that makes it safe
Claude is genuinely strong at this specific job for three reasons. It can read a messy export and understand the structure without you cleaning it first. It does not get tired or skip the boring cross-checks, so it actually looks at all the angles a rushed human skips. And it has no theory it is attached to, so it will not unconsciously steer the analysis toward the answer it wanted, which is the single biggest flaw in human diagnosis.
But there is one rule that makes the difference between useful analysis and confident nonsense.
Claude must only work from the data you give it. Never ask Claude what your conversion rate probably is, or what is normal for your industry, or to estimate a number it cannot see. It will produce a plausible-sounding figure, and it will be invented. The discipline is simple. Claude analyses the data in front of it. It does not recall, estimate, or assume. Every number in its output must be traceable to a number in your input. Build that into how you prompt it, and the analysis is trustworthy. Skip it, and you get fluent guesses dressed up as findings.
The workflow
Here is the actual diagnostic process, step by step.
Step one: get the data out. You do not need a data team. Export what you have as CSV. Funnel data from your analytics tool. The conversion or signup numbers broken down by whatever dimensions you can get, by week, by channel, by segment, by plan. The more granular the export, the better the diagnosis, because the cause of a moved number almost always lives in a breakdown, not in the aggregate.
Step two: give Claude the full context. Do not just paste numbers and ask what is wrong. Tell Claude what the business is, what the number is, what it normally does, what it did, and what the funnel steps are. Diagnosis without context produces generic answers. Diagnosis with context produces specific ones.
Step three, instruct Claude to segment before concluding. This is the most important instruction. The aggregate number that moved is never the real story. A conversion rate that dropped 12 percent overall did not drop 12 percent everywhere. It probably held steady in four segments and collapsed in one. The cause is in that one segment. Claude must break the moving number down by every dimension in the data before it offers any conclusion.
Step four, make it rank hypotheses by evidence, not plausibility. Once Claude has segmented the data, it should produce a ranked list of probable causes, and each one must be tied to the specific numbers that support it. A cause with strong evidence in the data ranks above a cause that merely sounds reasonable.
Step five, ask it what is missing. The data you exported cannot answer everything. A strong diagnostic ends with Claude telling you what additional data would confirm or kill the top hypothesis. That tells you exactly what to pull next instead of guessing again.
The operational prompt
Here is the complete prompt for diagnostic analysis. Paste it, attach your data export, fill in the bracketed context, and run it.
You are a senior growth analyst. I am going to give you a data
export and ask you to diagnose why a specific metric moved.
CRITICAL RULES:
- Work only from the data I provide. Do not estimate, recall,
or invent any number. Every figure in your analysis must be
traceable to my data.
- If the data cannot support a conclusion, say so explicitly.
Do not fill gaps with assumptions.
- Do not give me generic marketing advice. Every statement
must reference specific numbers from my export.
CONTEXT:
- Business: [what you do, business model, who the customer is]
- The metric: [which metric moved, e.g. trial-to-paid conversion]
- Normal behaviour: [what the metric usually is / its recent range]
- What happened: [what it did, over what time period]
- Funnel steps: [list the steps from entry to the metric]
- What I have already ruled out: [anything you know is not the cause]
YOUR TASK, in this order:
1. SEGMENT THE MOVE. Break the metric down by every dimension
available in my data (time, channel, segment, plan, geography,
device, whatever exists). Identify precisely where the move
is concentrated. State clearly: did the metric move evenly
everywhere, or did it move sharply in specific segments while
holding elsewhere?
2. ISOLATE THE LIKELY CAUSE. Based only on where the move is
concentrated, produce a ranked list of probable causes. For
each one, cite the specific numbers from my data that support
it. Rank by strength of evidence, not by plausibility.
3. STATE YOUR CONFIDENCE. For your top hypothesis, tell me how
confident the data lets you be, and what specifically in the
data drives that confidence.
4. NAME THE MISSING DATA. Tell me what additional data, if I
pulled it, would confirm or eliminate your top hypothesis.
Be specific about exactly what to export next.
5. SUMMARISE. In three sentences, plain language: what most
likely happened, where, and what I should investigate first.
Ask me clarifying questions before you begin if the context
I gave you is insufficient to do this well.The final line matters. It gives Claude permission to push back if you under-briefed it, which is what a good analyst does instead of producing confident analysis on bad inputs.
What this changes
The first time a team runs this properly, the reaction is usually the same. The real cause was not any of the theories that were being debated in the meeting.
That is the entire point. The conversion drop that the team had attributed to a pricing change turns out, once segmented, to be concentrated entirely in mobile traffic from one channel, which points to a broken checkout flow on a specific device, which is a completely different problem with a completely different fix. The team would have spent a quarter adjusting pricing, and the real problem would have kept running.
This is what diagnosis, instead of guessing, buys you. Not just speed, though, twenty minutes versus a never-completed afternoon of analysis is real. It buys you the correct problem. Acting fast on the wrong diagnosis is worse than acting slowly on the right one, and most teams, most of the time, are doing the first thing.
Run the next moved number through this workflow instead of through a meeting of competing theories. The cause is in the data. It almost always has been. The only thing that was missing was the hours to actually find it, and that is the constraint Claude removes.
Next edition, we move from diagnosis to foresight. Using Claude to find the problems that have not shown up in any number yet, the quiet decay in your cohorts that predicts the next quarter before it arrives.
See you at the next edition, Arindam