21 Aug 26 
AI makes it easy to slice audiences into hundreds of micro-segments, but over-segmentation quietly kills conversion. Here is what to personalise that actually matters, and how to keep relevance beating granularity.
CRO
Every few weeks a client asks me to build them a personalisation engine that speaks to hundreds of tiny audiences, and every time I have to slow the conversation down. AI has made slicing an audience almost free. You can now spin up a segment for left handed accountants in Pune who bounced twice on a Tuesday, and the tooling will happily let you. The trouble is that easy segmentation and better conversion are not the same thing. In my experience the studios that win at personalisation are the ones that resist most of the slicing they could do, and spend that energy on relevance instead. Granularity feels like progress. Relevance is progress.
Why more segments is not more conversion
The maths of over-segmentation is quietly brutal. Split a page that gets 10,000 monthly visitors into forty micro-segments and you have handed each variant a couple of hundred people, most of whom will never convert. You can no longer tell whether a variant works or whether you are staring at noise. Thin data per segment means slow tests, wide confidence intervals, and decisions made on vibes. A single strong message tested against real volume will usually beat a dozen clever ones that never reach significance.
There is a second cost that never shows up in the pitch deck: maintenance. Every variant is a thing someone has to write, design, QA, and keep truthful as the offer changes. Forty variants is forty chances for a broken price, a stale headline, or a promise the product no longer keeps. I have watched teams spend more hours babysitting their segmentation than they ever recovered in lift.
Then there is the creepiness tax. When personalisation gets too specific, people feel watched rather than served. Naming someone’s employer, referencing a page they lingered on, guessing at something private: these read as surveillance, and they cost trust faster than any tuned headline can earn it back. Relevance reassures. Granularity, past a point, unsettles.
What actually deserves personalising
Most of the value lives in a handful of broad, high signal cuts. Intent and stage first: a first time visitor and a returning cart abandoner are genuinely different people with different questions, and treating them the same is the real waste. Source and context next, because someone arriving from a price comparison search wants different reassurance than someone clicking a brand campaign. After that, a few durable attributes if you truly have them cleanly, ideally declared rather than inferred, which is why I keep coming back to a zero-party data approach where people tell you what they want.
Everything beyond that tends to be vanity slicing. Personalising by device when the message would not change, by weather, by a demographic guess you cannot verify, none of it moves the number. The same discipline that keeps email segmentation sane applies on site: fewer, meaningful groups beat many hollow ones. Before you personalise anything, run a conversion friction audit first, because a clearer single message often outperforms a personalised muddle.
Using AI responsibly, and knowing if it worked
Here is the workflow I trust. Use AI to find broad useful segments and to draft the variants, then test each one against a strong control before rollout, never straight to production. Put guardrails against creepiness in writing: no naming employers, no referencing private behaviour, a tone that helps rather than reveals. And always build a fallback, a solid default message that renders when the data is missing, because it will be missing more often than anyone admits.
Measurement is where honesty lives. Judge personalisation net of its complexity. If a segment lifts conversion by two percent but doubles your maintenance load, it lost. Watch micro-conversions as leading indicators to read direction early, and hold segments to real significance, not hope.
A short checklist I give clients: personalise by intent and stage before anything else; keep segments broad enough to reach significance; test every variant against a strong single message; write your creepiness guardrails down; always ship a fallback; and retire any segment that costs more to maintain than it returns. Relevance beats granularity, every time.