About

Most of this field
has never been tested twice.

Which is a strange thing to say about an industry with this many confident opinions in it. It is also the entire reason this agency exists.

Chief Recommending Officer

Abhishek Mishra

Chief Recommending Officer

50+

Brands helped to land in AI answers across ChatGPT, Gemini, Perplexity and Claude.

I have worked on answer engine optimization since the discipline had a name — testing tactics while the surfaces themselves kept changing underneath, back when most of the field was still arguing about whether any of this was real.

That period taught me something uncomfortable: a large share of what gets published as AEO best practice has never been tested against a control group, and a fair amount of it contradicts the rest. Two well-regarded agencies will teach opposite playbooks for the same channel in the same month, and neither will show you the data.

So the method here is duller than the marketing in this industry: run the experiment, hold a control group aside, watch most tactics fail, keep the few that reproduce more than once. Publish the failures alongside the wins, because a field this young is better served by traceable sources than by another confident voice.

When something is genuinely unsettled — and plenty is — you will hear that from me rather than a clean answer I cannot support.

Method

How a tactic earns its place.

This is the loop every client programme runs on, and the same one behind everything we publish.

Build the question set

Reconstructed from your sales call transcripts, support tickets, competitor paid-search keywords converted into natural questions, and the threads your buyers already post in. Not a keyword list with question marks bolted on.

Take a baseline, across surfaces

Each question asked repeatedly and in several phrasings, because an answer is a weighted sample rather than a fixed result. ChatGPT, Gemini, Perplexity and Claude tracked separately, since their citation behaviour genuinely differs.

Split into test and control

One tactic per test group, one group left deliberately untouched. Without the control you cannot tell your work apart from a platform update, and platform updates move these numbers more than most people admit.

Run it long enough to mean something

A few weeks, not a few days. New content can spike briefly and then decay, so an early reading will flatter almost anything you do.

Reproduce before believing

A tactic counts once the test group moved, the control did not, and the result held up on a second run. Anything that only worked once goes back in the queue, not into your strategy.

Boundaries

What we will not do.

Some of these will cost us work. They are still the right side of the line, and the downside of the alternative is not a zero — it is a reputation you then have to spend money repairing.

Fake community accounts

No sock puppets, no undisclosed personas, no seeded threads. It gets detected, the comments get removed, and the account that carried your brand gets banned with your name on it.

Fully automated content

AI-assisted with a human editing is normal and fine. Fully generated at volume correlates with worse citation performance, not better. We will not sell you more of something the data says underperforms.

Guaranteed placement

Nobody outside the model providers controls what an answer says. Anyone promising a guaranteed spot in ChatGPT is either misunderstanding the system or counting on you to.

Numbers without a source

No statistic goes in a deck of ours without a named methodology behind it. If we cannot tell you the sample and the method, you should not be making decisions on it.

Start

An hour is a low-risk way to find out.

rankme@recommendrai.com