AI search conversion benchmarks

The conversion spread

Why the numbers disagree so much

• "Conversion" isn't one definition. Some datasets count any goal completion; others count only revenue events.
• The baseline differs. Some compare AI referrals to all organic traffic; others to non-brand organic specifically — a meaningfully different comparison.
• AI referral volume is tiny — around 1% of traffic per large panels — so small absolute conversion counts produce unstable ratios that swing wildly between studies and periods.

What "conversion" actually means in each study

This is the root of the problem, and it is worth being concrete about. "Conversion rate" sounds like a standard metric. It is not. It is a family of metrics that share a name.

Any goal completion. The broadest definition. A newsletter signup, a PDF download, a demo request, and a purchase all count equally. Sites using this definition report much higher rates than sites using a narrower one, for reasons that have nothing to do with traffic quality.

Revenue events only. A completed purchase, or a qualified opportunity created. Much stricter, and produces rates several times lower on the same traffic.

Session-level vs. user-level. Does a visitor who converts on their third session count once, or does the rate get computed per session? For a considered purchase with a long research cycle, this choice alone can move a reported rate substantially.

First-touch vs. last-touch attribution. If someone discovers you through an AI answer, leaves, and returns via a branded Google search a week later, which channel gets the conversion? Almost every analytics setup answers this differently, and almost no published study states which it used.

Now look back at the spread table. A study using "any goal completion, session-level" against a study using "revenue only, user-level, last-touch" could report wildly different multiples while both measure honestly and competently. The disagreement may not indicate anyone is wrong. It may indicate they are measuring different things and calling them the same.

This is why the honest recommendation is to ignore the multiples entirely and measure your own. Not because outside numbers are dishonest, but because a multiple built on someone else's definition cannot be transferred to yours.

The baseline problem, in detail

The second half of any multiple is what it compares against, and that half gets even less scrutiny than the definition of conversion itself.

All organic traffic as baseline. This includes branded search: people typing your company name, who already know who you are and convert at very high rates. Including them raises the baseline, which shrinks the apparent AI multiple.

Non-brand organic as baseline. This excludes branded search and is the fairer comparison for most purposes, since AI referrals are typically discovery traffic rather than people who already knew you. Excluding brand lowers the baseline, which inflates the apparent multiple.

That single choice can move a reported multiple by a large factor without anything changing about the AI traffic at all. Two studies could observe identical AI-referral behaviour and report very different multiples purely from this decision.

There is a third baseline nobody uses and arguably should: comparing AI referrals against non-brand organic traffic on the same queries. That would control for topic and intent together. It is harder to assemble, which is presumably why it does not appear in any published figure we found.

Why small volume breaks these ratios

The third structural problem is statistical, and it gets almost no attention in coverage of these numbers.

AI referral traffic remains a small share of most sites' total, around one percent by one large panel's measure. Small session counts produce small conversion counts. Small conversion counts produce unstable rates.

Consider the arithmetic. A segment with 200 sessions and 4 conversions reports a 2% rate. One more conversion, a single visitor, moves it to 2.5%: a 25% relative swing from one person's behaviour. Compare that against an organic segment with 40,000 sessions, where one extra conversion changes nothing meaningful.

Now compute a multiple from those two. The numerator swings wildly, the denominator barely moves, and the resulting ratio inherits all of the instability. Publish it as a headline figure and it reads as solid. It is not.

This also explains something otherwise puzzling about the spread: why the reported multiples vary so much more than the underlying conversion rates do. Ratios of a noisy number to a stable one are much noisier than either input.

The confound nobody controls for

Here's the likely biggest uncontrolled factor. Sites that get cited in AI answers may already skew toward considered, higher-intent purchases and well-established brands. That's true regardless of any effect AI referral traffic itself has.

If so, most of the reported "AI traffic converts better" effect may just reflect which sites get cited. It may say nothing about what AI referral traffic actually does to a site's conversion rate. No published study we found controls for this.

Three more confounds worth naming

The site-selection confound above is the biggest one. It is not the only one, and the others point in different directions, which matters.

Query intent. People turn to AI assistants disproportionately for research and comparison questions. Those queries sit further down the consideration funnel than a generic informational search. If AI referrals arrive from more decision-ready questions on average, higher conversion follows from intent, not from the channel.

Pre-qualification by the answer itself. An AI answer often summarises what a product does before the visitor arrives. Someone who reads that summary and still clicks through has effectively self-selected. The people who would have bounced immediately never became a session at all. That removes low-intent visitors from the denominator, which raises the rate without anything about the traffic being inherently better.

Undercounted volume, in the other direction. Much AI-sourced traffic likely arrives with no detectable referrer and lands in "direct." If that hidden traffic converts worse than the visible AI-referral segment, every published multiple is computed on a favourably biased sample. This is the one confound that could mean the true effect is smaller than reported rather than larger.

Taken together, these suggest the direction of the finding is probably real, and the magnitude is almost certainly overstated. Two of the three confounds inflate the number. The third means we may not even be measuring the whole population.

How you would actually test the confound

A cleaner test compares AI-referral conversion rate against organic conversion rate within the same site. That holds the site's own baseline conversion tendency constant. It beats comparing across a mix of different sites with different baseline rates.

This isn't a study that needs new data collection. Most sites with any AI referral volume could run this comparison on their own analytics today. It's also a more useful benchmark for one specific business than any aggregate multiple in the table above.

Why AI traffic might genuinely convert better

Everything above is deflationary, so it is worth stating the case for the effect being real. The mechanisms are plausible, and dismissing the finding entirely would be as unearned as accepting the headline multiple.

Implicit endorsement. Being named in an answer to someone's question is closer to a recommendation than a link in a ranked list. A visitor arriving from that context has been told, in effect, that a system found you relevant enough to cite. That framing plausibly carries more trust than a blue link the reader picked from ten options.

Better-informed arrival. The visitor has already read a summary of what you do. They are unlikely to arrive with a fundamental misunderstanding of your product, which is a common source of immediate bounces from ordinary search.

Deeper questions. People ask assistants longer, more specific, more considered questions than they type into a search box. Specific questions correlate with being further along in a decision.

Reported engagement. Panel data describes AI-referred visitors spending notably more time on site and viewing more pages per visit than average. Engagement is not conversion, but it is consistent with visitors who arrived for a reason rather than by accident.

So the reasonable read is not "the effect is fake." It is that a real effect and several inflating confounds are being reported as one number, and nobody has separated them. That is a genuinely different claim from either "AI traffic converts 4.4x better" or "the finding is meaningless."

The honest version of this statistic

AI referral traffic appears to convert better than average organic traffic, by an amount nobody has credibly established with a consistent definition and a controlled comparison. That sentence is less quotable than "4.4x higher conversion" — and considerably more accurate.

Three things are worth separating in that sentence. The direction is reasonably well supported: multiple independent sources point the same way. The magnitude is not supported at all: the published range spans an order of magnitude. And the causal story, that the channel itself produces the lift, is unsupported, because the confounds above have never been controlled for.

Most coverage collapses those three into one confident claim. Keeping them apart is what makes the statement defensible when someone asks where the number came from.

How to benchmark your own AI traffic

Segment your own analytics by AI-platform referrer, define "conversion" once and apply it consistently to both your AI-referral and organic segments, and compare against your own non-brand organic baseline rather than an external multiple built on a different definition entirely.

Concretely, that means four decisions made before you look at any numbers. Decide which referrer patterns count as AI traffic, and write the list down. Decide what a conversion is, in one sentence. Decide whether your baseline includes branded search, and exclude it unless you have a specific reason not to. Decide the window, long enough to accumulate a workable number of conversions in the smaller segment.

Making those calls in advance matters more than which way you decide them. Any consistent definition produces a usable trend over time. A definition chosen after seeing the data produces a number that flatters whatever you hoped to find, and you will not be able to tell the difference later.

Then re-run the same comparison each quarter without changing any of the four decisions. The absolute multiple matters far less than whether it is moving, and only a frozen method can tell you that.

A worked benchmarking example

Here's the method with invented numbers. Say a hypothetical site logs 400 AI-referral sessions over a quarter, with 18 conversions. That's a 4.5% rate. Over the same period, it logs 40,000 non-brand organic sessions with 800 conversions, a 2.0% rate. Same conversion definition, both segments.

That works out to AI referral traffic converting at roughly 2.25× this site's own organic baseline. It's imprecise, since the AI-referral sample is small. But it's real and defensible for this specific business. That makes it far more useful for a budget conversation than any externally published multiple.

Measurement pitfalls in your own data

Measuring your own site avoids the comparability problem. It introduces a few of its own. These are the ones that most often produce a number that looks meaningful and is not.

Reading a rate off a tiny sample. The single most common error. If your AI-referral segment has produced eight conversions this quarter, you do not have a conversion rate. You have eight events. Extend the window until you have enough conversions per segment that one more would not visibly move the number.

Changing the conversion definition mid-comparison. Counting demo requests for one segment and closed revenue for the other produces a difference that says nothing about the traffic. Fix the definition first, apply it to both, then compare.

Leaving branded search in the baseline. Your branded organic traffic converts far better than anything else and is not comparable to discovery traffic. Leaving it in makes your AI multiple look smaller than it should. Most people do this accidentally.

Ignoring the direct bucket. Some AI-sourced visits arrive with no referrer. You cannot fully solve this, but you can watch for it: a rise in direct traffic that coincides with rising AI citations is a signal that your measured AI segment understates the real thing.

Comparing across a period where something else changed. A pricing change, a campaign, or a seasonal peak inside your measurement window will contaminate both segments unevenly. Prefer a window where nothing structural moved, even if it means waiting.

Reporting a ratio without the underlying counts. "2.25x" invites confidence. "18 conversions from 400 sessions, versus 800 from 40,000" invites appropriate caution. Always publish both internally, whatever you lead with.

The attribution problem underneath all of this

There is a deeper issue that limits every number on this page, including the ones you measure yourself. Standard analytics was built for a world of clicks, and a meaningful share of AI-sourced value never produces one.

Consider four ways someone can encounter you through an AI system. They click a citation, arriving with a detectable referrer: this is the only case your analytics sees cleanly. They read your content summarised inside the answer and never click at all. They see your brand named, then search for you directly a few days later, arriving as branded organic. Or they reach you through an interface that strips the referrer entirely, landing in direct.

Only the first shows up as AI referral traffic. The second produces value you cannot measure at all. The third and fourth produce value your analytics attributes to a different channel entirely.

That has a specific consequence for conversion benchmarking. Your AI-referral segment is not a random sample of people who encountered you via AI. It is the subset who both clicked and arrived with an intact referrer. Whether that subset converts like the whole population is unknown and unknowable with current tooling.

This is not a reason to stop measuring. It is a reason to treat your own multiple as a directional signal about one visible segment, rather than a complete account of what AI search is doing for your business. Full treatment of the undercounting problem is on the referral traffic page.

How to report this to a stakeholder

The practical question most readers actually have: what do you put in the deck? A few principles that keep a claim defensible if someone pushes back.

Lead with your own number, not a published multiple. "Our AI-referral traffic converted at 4.5% against a 2.0% non-brand organic baseline last quarter" is checkable and specific to the business. "Industry data shows AI traffic converts 4.4x better" is neither.

State the counts alongside the rate. This preempts the obvious challenge and shows you know the sample is small. A stakeholder who discovers the small sample themselves will discount everything else you said.

Give the direction more weight than the magnitude. "This segment converts better than our organic baseline, on a small sample" is a claim that will hold up. A precise multiple from a quarter's worth of AI traffic will not survive re-measurement, and being wrong next quarter costs more credibility than being vague this quarter.

Name the volume constraint explicitly. If AI referrals are one percent of sessions, a high conversion rate on that one percent is not yet a material revenue line. Say so before someone else does. It protects the argument that this is a growing channel worth early investment, rather than overselling it as a current one.

Frame it as a trend to watch, not a result to bank. The defensible version of this argument is that the channel is small, appears to convert well, and is growing, so it warrants attention now. That framing survives contact with the data. A specific multiple does not.

What would actually settle this

It is worth being specific about what a credible answer would require, because "more research" is not a useful ask. Four things would move this category from broken to something usable.

A published definition alongside every figure. The cheapest fix available, and it needs no new data collection. If each vendor stated what counted as a conversion, what the baseline was, and whether attribution was first- or last-touch, several of the existing numbers might reconcile immediately. The spread might partly be a documentation failure rather than a measurement one.

Within-site comparisons, aggregated. Rather than comparing AI-referral conversion across a mix of different sites, collect the AI-versus-organic ratio computed within each site, then look at the distribution of those ratios. That controls for the site-selection confound directly, because each site serves as its own control.

Segmentation by query intent. Splitting results by whether the originating question was research-stage or decision-stage would separate the intent effect from any channel effect. Nobody publishes this, and it is probably the single most informative cut available.

An honest account of the invisible traffic. Any study that measures only visible AI referrals should say what share of AI-sourced visits it believes it is missing. Even a rough estimate would tell readers how much of the population the finding actually covers.

None of those require a breakthrough. They require someone measuring carefully and publishing the method. That this has not happened, in a field with substantial commercial interest in the answer, is itself informative about how the field's incentives run.

Verification status

Every individual multiple above traces to a named source, but none can be reconciled with the others — this entire category is graded Broken chain as a set, per the provenance audit, even though each individual number is a real, attributed figure.

Worth being precise about what that grade means, since it is unusual. Broken here does not mean any individual study was done badly, or that any specific figure is false. Each one traces to a named source with a real sample. The grade describes the set: four attributed numbers that cannot be reconciled with each other, published without the definitions that would let anyone reconcile them.

A reader encountering any one of these figures in isolation would have no way to know the others exist, or that they disagree by an order of magnitude. Documenting that is the entire purpose of grading a category rather than only individual claims.