AI search conversion benchmarks

How far apart are the published figures?

The 23× claim sits at the top of that range and is the one most often repeated second-hand, usually without the caveat that neither its definition of conversion nor its sample is disclosed. The category's row in the master statistics hub carries the same grade this page ends on.

Why the four figures cannot be compared

Set them side by side with what each one actually measured, and the spread stops looking like a contradiction and starts looking like four different metrics sharing a name.

Three structural differences explain the spread without anyone having measured incompetently: what counts as a conversion, what the multiple is measured against, and how few conversions sit under the AI-side rate. Each gets a section below.

What "conversion" means in each study

"Conversion rate" sounds like a standard metric. It is a family of metrics that share a name, and the published figures above do not say which member of the family they used.

• Any goal completion. Newsletter signup, PDF download, demo request and purchase all count equally. Produces much higher rates than a narrower definition, for reasons unrelated to traffic quality.
• Revenue events only. A completed purchase, or a qualified opportunity. Several times lower on identical traffic.
• Session-level vs. user-level. Does a visitor converting on their third session count once, or is the rate computed per session? On a considered purchase with a long research cycle, this choice alone moves the rate substantially.
• First-touch vs. last-touch. Someone discovers you in an AI answer, leaves, returns via branded search a week later — which channel books the conversion? Almost every analytics setup answers differently, and almost no published study states which it used.

A study using "any goal completion, session-level" and a study using "revenue only, user-level, last-touch" could report wildly different multiples while both measure honestly. The disagreement need not indicate anyone is wrong. It indicates they are measuring different things and calling them the same — which is why a multiple built on someone else's definition cannot be transferred to yours.

Which baseline the multiple is measured against

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

• All organic traffic as baseline includes branded search — people typing your company name, who already know you and convert at very high rates. That raises the baseline and shrinks the apparent AI multiple.
• Non-brand organic as baseline excludes them, and is the fairer comparison for most purposes, since AI referrals are typically discovery traffic. That lowers the baseline and inflates the apparent multiple.
• Non-brand organic on the same queries would control for topic and intent together. Nobody uses it; it is harder to assemble, which is presumably why it appears in no published figure we found.

Two studies could observe identical AI-referral behaviour and report very different multiples purely from that decision. None of the four figures states which it made.

Why small volume breaks these ratios

AI referral traffic remains a small share of most sites' total — around one percent by Similarweb's generative-AI usage panel, consistent with the figures collected on the market share page. Small session counts produce small conversion counts, and small conversion counts produce unstable rates.

The arithmetic is stark. A segment with 200 sessions and 4 conversions reports 2%. One more conversion — one person — moves it to 2.5%: a 25% relative swing. The organic segment it is divided by has 40,000 sessions, where one extra conversion changes nothing. The numerator swings, the denominator barely moves, and the ratio inherits all of the instability. That is why the reported multiples vary far more than the underlying conversion rates do.

The confounds nobody controls for

Even a perfectly consistent definition would leave four uncontrolled factors, and they do not all point the same way.

Site selection — the largest. Sites that get cited in AI answers may already skew toward considered, higher-intent purchases and well-established brands, regardless of any effect the referral channel has. If so, much of the reported effect reflects which sites get cited rather than what AI referral traffic does. No published study we found controls for this.

Query intent. People turn to assistants disproportionately for research and comparison questions, which sit further down the consideration funnel. Higher conversion would then follow from intent, not from the channel.

Pre-qualification by the answer. An AI answer often summarises what a product does before the visitor arrives. People who would have bounced immediately never become a session at all, which removes low-intent visitors from the denominator and raises the rate without the traffic being inherently better.

Undercounted volume, pointing the other way. Much AI-sourced traffic likely arrives with no detectable referrer and lands in "direct". If that hidden traffic converts worse than the visible 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.

Hypothesis Two of these inflate the number and one means we may not be measuring the whole population — so the direction is probably real and the magnitude is probably overstated. A cleaner test would compare AI-referral against organic conversion within the same site, holding that site's own baseline tendency constant. It needs no new data collection: most sites with any AI referral volume could run it on their existing analytics today.

Why AI traffic might genuinely convert better

Everything above is deflationary, so the case for a real effect is worth stating; dismissing the finding entirely would be as unearned as accepting the headline multiple.

• Implicit endorsement. Being named in an answer is closer to a recommendation than a link in a ranked list, and that framing plausibly carries more trust than a blue link picked from ten options.
• Better-informed arrival. The visitor has already read a summary of what you do, so they are unlikely to arrive with the fundamental misunderstanding that causes many immediate bounces from ordinary search.
• Deeper questions. People ask assistants longer, more specific questions than they type into a search box, and specificity correlates with being further along in a decision.
• Reported engagement. Panel data describes AI-referred visitors spending 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. Ahrefs' AI SEO statistics compilation reports the same directional pattern, with the same sample-size caveat.

Scale, at least, is not in doubt. Statista's monthly ChatGPT user series tracks an audience large enough that even a one-percent referral share is a material number of sessions for many sites. The channel being small today is a statement about attribution and click behaviour, not about how many people are asking the questions.

Do the platforms differ from each other?

The per-engine row is the most defensible in the table, because it is a within-study comparison: one vendor, one definition held constant across three engines, one window. It is still vendor-reported, and its population is undisclosed.

A per-engine difference would not be surprising, because the engines differ upstream of conversion in documented ways: they cite different kinds of sources, as Discovered Labs' side-by-side comparison and Profound's platform citation patterns both describe, and they surface links with different prominence. Somebody arriving from a cited forum thread and somebody arriving from a cited product page are not the same visitor, and averaging them into one channel rate hides that. Our own per-engine collections sit on ChatGPT, Perplexity and Claude citation statistics; the last is the shortest by a wide margin, and the absence is the finding.

Hypothesis That per-engine conversion differences are real and stable. Nothing here establishes it. One vendor study measured it once, and the concordance work suggests the engines overlap less than most people assume — which would make a blended figure even less meaningful than a blended citation figure.

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 is less quotable than "4.4x higher conversion" and considerably more accurate.

Three things stay separate 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 a fivefold gap. 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 the three into one confident claim.

Where the benchmark probably differs by sector

Every published multiple is a cross-industry aggregate, and there are specific reasons to expect large variance underneath.

• Considered B2B purchases. Long cycles and multiple sessions per buyer, with a conversion event that may be a demo request rather than revenue. Session-level rates flatter this category badly.
• Retail and commerce. Shorter cycles and a cleaner revenue event, but an increasingly agent-mediated path — see the AI shopping statistics. When the visitor is partly software, "conversion rate" measures something the metric was not designed for.
• Local and service businesses. The conversion event often happens off-site entirely: a phone call, a walk-in, a third-party booking. This is also the sector with the thinnest research base, which the local AI search page states plainly rather than papering over.
• Health, finance and other YMYL categories. Engines behave differently where the cost of a wrong answer is high, changing both which sources get cited and how much of the answer is given without a link.

Open question No published dataset breaks conversion down by sector at all. Until one does, treating a cross-industry multiple as applicable to your vertical is an assumption, not a benchmark.

How to benchmark your own AI traffic

Make four decisions before you look at any numbers, then hold them fixed. Which way you decide them matters less than deciding in advance: any consistent definition produces a usable trend, while a definition chosen after seeing the data flatters whatever you hoped to find, and you will not be able to tell the difference later.

A worked example, with invented figures. A hypothetical site logs 400 AI-referral sessions over a quarter with 18 conversions — a 4.5% rate — against 40,000 non-brand organic sessions with 800 conversions, a 2.0% rate, same conversion definition on both. That is roughly 2.25× this site's own organic baseline: imprecise, because 18 conversions is a thin sample, but real and defensible for this specific business, and far more useful in a budget conversation than any published multiple. Note how much more confident "2.25×" reads than "18 conversions from 400 sessions". Always report the counts beside the rate — internally and externally.

Four pitfalls account for most bad self-measurements: reading a rate off a handful of conversions; changing the conversion definition between segments; leaving branded search in the baseline, which makes your multiple look smaller than it is; and choosing a window in which a pricing change, campaign or seasonal peak contaminated both segments unevenly.

Why your own number is also a partial view

Standard analytics was built for a world of clicks, and a meaningful share of AI-sourced value never produces one. There are four ways someone can encounter you through an AI system, and only the first is visible.

The third path — brand recalled, searched directly days later — is the most often overlooked. RankScience's write-up of the brand-mention visibility gap describes the same shape: being named in an answer produces downstream effect that never appears as a referral.

The consequence for benchmarking is precise. 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 with current tooling. Treat your own multiple as a directional signal about one visible segment. The undercounting problem is covered in full on the referral traffic page, and how often no click happens at all on the zero-click page and the CTR page.

What would actually settle this

"More research" is not a useful ask. Four specific things would move this category from broken to usable, and none requires a breakthrough.

• A published definition beside every figure. The cheapest fix, needing no new collection. Six fields would do it: what counted as a conversion, in one sentence; session-level or user-level; first-touch or last-touch; what the baseline segment was and whether branded search was excluded; the absolute counts, not only the rate; and the window, with dates. Any figure carrying those six is comparable to any other that does. Part of the spread may be a documentation failure rather than a measurement one.
• Within-site comparisons, aggregated. Collect the AI-versus-organic ratio computed inside each site, then look at the distribution. Each site serves as its own control, which addresses the site-selection confound directly.
• Segmentation by query intent. Splitting by whether the originating question was research-stage or decision-stage would separate the intent effect from any channel effect. Probably the single most informative cut available, and nobody publishes it.
• An honest account of the invisible traffic. Any study measuring only visible AI referrals should state what share of AI-sourced visits it believes it is missing.

That none of this has happened, in a field with substantial commercial interest in the answer, is itself informative. The same argument in general form is on the measurement standard page, and our own attempt at supplying the missing measurement layer is the AI Citation Index, with negative findings going to the null results registry.

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.

Broken here does not mean any individual study was done badly, or that any specific figure is false. Each traces to a named source. 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 them in isolation would have no way to know the others exist, or that they disagree by a factor of five. Documenting that is the entire purpose of grading a category rather than only individual claims.