Dark Funnel Metrics: What to Measure When Buyers Decide Before They Click

Picture a marketing director asking ChatGPT which competitive intelligence tools to shortlist for her team. Your brand is mentioned in the answer, but there’s no link, so there’s no click, no session, and no evidence recorded for that search.

Four days later she searches your name on Google, lands on your pricing page, and leaves. Your analytics counts one organic visit from a branded query.

The last step is logged correctly. The cause is not logged at all.

That gap is why dark funnel metrics now have to do the job traffic used to.

Gartner’s survey of 632 B2B buyers found that 61% prefer a rep-free buying experience. G2, a software review marketplace, surveyed 1,076 software buyers in March 2026 and found that 51% now start their research in an AI chatbot more often than in Google, and that 69% chose a different vendor than planned because a chatbot recommended it. The decision forms in places that never send you a session.

Traffic records the visit after the decision, making it a late witness (one that insists it saw everything).

Here is my thesis: in a dark funnel you cannot measure the activity. You can measure its footprint: how often AI answers name you, which of your pages they cite, how branded demand moves afterward, and what buyers say when you ask them.

Each one is a proxy, and proxies lie. So each one needs its own denominator, a validation step before you let it move budgets, and a firm rule: never add them together.

This is the measurement follow-up to my article on marketing funnel fragmentation, which explained why the journey splits across systems that cannot be matched. Here you’ll get a working definition, the five metrics that replace traffic as the verdict, the LEAD framework to decide which proxies to trust, and a guide to which metric to lead with for your business model.

What is a dark funnel metric?

A dark funnel metric measures buyer influence without click evidence. It tracks either upstream presence, such as how often AI answers name your brand, or downstream effects, such as how branded demand moves and what buyers say brought them. It is a proxy for influence, not a count of visits.

What is the difference between a direct metric and a proxy metric?

A direct metric counts an event you observed for a specific person: a session, a form fill, a purchase. A proxy metric measures a related signal in aggregate and infers influence from it: branded search volume, share of AI mentions, the percentage of buyers who name a podcast. Every dark funnel metric is a proxy.

That is not a flaw to apologize for. It is the price of measuring something nobody recorded. It does mean every proxy needs two things a session never did: a defensible denominator, and a test that it tracks something real. Both are covered below.

How does a dark funnel metric relate to dark social and funnel fragmentation?

Dark social, a term Alexis Madrigal coined in The Atlantic in 2012, is the private sharing of links through messages, email forwards, and chat apps, which strips the referrer and files the visit under direct. The dark funnel is the wider set: dark social plus podcasts, private communities, event conversations, and AI answers. Funnel fragmentation is a different problem: every step is recorded, but never as one person.

Roi Kaufman’s guide to the B2B dark funnel covers the channels and the demand generation playbook. The metrics work for both problems because neither requires matching one user across systems. They operate at the group level, where connections do exist.

How much of the dark funnel can you actually measure?

Only the footprint, not the activity. You can measure that footprint at four layers, and your confidence drops as you move down the table.

Footprints can be measured in aggregate.

Similarweb’s Downstream Impact of AI Visibility study followed a panel’s clickstream across sites for seven days after a ChatGPT brand recommendation (US desktop users, July to December 2025) and found those users were 2.5 times more likely to visit the recommended brand than a direct competitor. It sized the effect without analyzing a single conversation.

How much of the journey is dark, as a percentage? Nobody knows. Published estimates use different definitions of “invisible,” so I do not quote one. A number you cannot reproduce is a worse planning input than an honest range built from your own self-reported attribution and branded search data.

Which traffic KPIs should you retire, and what replaces each?

Retire sessions from AI referrers, organic sessions per page, direct traffic growth, and last-touch source as verdicts on performance. Replace them with share of AI mentions, citations by URL, branded search lift, self-reported attribution, and unexplained direct share.

Keep traffic as a diagnostic: it tells you where people land, not whether your marketing worked.

Kevin Indig, founder of Growth Memo, made the same point in Similarweb’s 2025 Generative AI Landscape: “AI traffic is not the leading metric in this era.”

“Retire” means retire as a judge, not delete from the dashboard. Three data points explain why, and the first two are covered in detail in the fragmentation article.

• AI influence doesn’t show up as AI traffic: In the Similarweb study, 55.9% of AI-influenced visits came through search, 19.9% through direct, 13.5% through other referrals, and 8.8% through an AI referral. Standard visits split 40.4% search and 38.8% direct. A report that sizes AI from referral data captures roughly a tenth of the effect.
• AI sits in the middle of the journey: In Similarweb’s State of Ecommerce 2026 report (US desktop shopper journeys, June 2026), AI was a middle touchpoint in 76.1% of journeys where it appeared, compared with 23.2% as the first touchpoint and 18.4% as the last. First-touch models miss it about 3 out of 4 times. Last-touch models miss it more than 4 out of 5 times.
• Presence in an answer no longer implies a visit: The State of Ecommerce 2026 report found that gen AI platforms send just 0.4% of retail ecommerce visits. On the search side, SparkToro’s analysis of Similarweb clickstream data found that 68.01% of US Google searches ended without a click between January and April 2026, up from 60.45% in 2024. That is the click side of zero-click marketing.

None of this is new in spirit. In 2022, Rand Fishkin estimated that 95% of hard-to-measure marketing shows up in three places in your analytics: direct visits, branded search, and lifts in click-through and conversion rates. His fix was a correlation dashboard, and in 2026 he is still telling marketers to replace traffic as a KPI with one.

He was right. What changed is that a fourth footprint now sits upstream of all three, and unlike a podcast or a Slack thread, it can be tracked: the AI answer itself.

Here is the retire-and-replace map.

Every replacement carries its own denominator. Share of mentions and branded search lift measure different populations, so report them side by side and read the pattern between them.

This is the same rule behind the METER framework in the fragmentation article, applied to what you stop reporting.

One more rule: the replacements are candidates, not proven causes. Each has to pass the LEAD check for your brand before it earns a place in your report.

How do you calculate the five dark funnel metrics?

Each metric needs four things: a numerator, a denominator, a time window, and a named source. Without all four, you can’t compare month to month or trust a trend line built on it. Three of the five metrics come from Similarweb. The other two come from your own CRM, because no panel can see what a buyer typed into your form.

In Similarweb, the AI Brand Visibility tool in the AI Search Intelligence suite tracks mentions and citations by prompt set, topic, and engine, and the keyword research tool (inside the SEO suite) supplies branded search data. Self-reported attribution and unexplained direct share live in your CRM and analytics.

How do you calculate the share of AI mentions?

Share of AI mentions is the percentage of AI responses in your category that name your brand. Divide the responses that mention you by all tracked responses, then split by engine, because each engine behaves differently.

The number only means something against a stable prompt set and a category denominator. The AI share of voice guide covers the full method, and it provides industry benchmarks for what a good AI brand visibility score looks like.

Similarweb reports two versions of this metric:

1. Visibility is responses mentioning you ÷ all tracked responses.
2. Mention share is your mentions ÷ all brand mentions across those responses.

In the same period, they can move by different amounts, because the totals underneath them change differently. Pick one, write down which, and freeze the prompt set. To read direction instead of level, AI Visibility Momentum indexes the mention rate against a baseline.

The denominator also tells you how this metric relates to branded demand. In Similarweb’s 2026 AI Brand Visibility Index (113 brands across 6 industries, US, January 2026), specialist and education-led brands in every sector ranked significantly higher in AI visibility than their branded search demand would suggest.

AI visibility and branded search demand measure different things, so one can rise while the other stays flat. That is why this article never measures them against the same denominator.

How do you measure citations by URL?

The Citations by URL metric counts how often each of your pages is used as an AI citation, mapped to folder depth. Split it from traffic, because the page that earns the citation is usually not the page that receives the visit.

Similarweb’s 2026 GenAI Landscape report compared ChatGPT citations with ChatGPT referral traffic (US, desktop, May 2026). Among cited URLs, 41.7% sat at folder depth two. Among referral visits, 58.8% landed on the homepage.

Score proof pages on citations and influence score, and arrival pages on conversion readiness. A comparison page with 40 citations and zero visits is not underperforming. It is doing its job.

You can see the same split in your own category before you build anything. Take the beauty category, for example: on Similarweb’s AI visibility leaderboard for beauty, the most-cited domain in ChatGPT answers for September 2026 (US) was the American Academy of Dermatology, followed by Allure, with Ulta and Sephora in third and fourth.

Four of the five most-cited pages were the Academy’s skin care and acne guides, and the only retailer page in the top five was a Sephora returns policy. The pages AI answers lean on are not the pages shoppers buy from.

How do you calculate branded search lift?

Branded search lift is the change in branded search clicks in a fixed window after a visibility event. Compare the seven days after an AI visibility change, a press hit, or a podcast with the seven days before, then divide by your own baseline: (branded clicks after ÷ branded clicks before) minus 1. It is the closest observable echo of influence that never produced a click.

The seven-day window comes from the Similarweb downstream study mentioned before, which tracked visits for seven days after a ChatGPT recommendation. Treat it as a starting window, not a law, and measure your own lag across several events before you hard-code it.

Segment your branded queries before you calculate anything.

• Navigational terms (login, sign in, mail, support) come mostly from people who already own the product.
• Evaluative terms (product names, “what is,” “vs,” pricing, reviews) come from people deciding.

A lift calculated on the blended head term mixes the two, and a small change in evaluative demand can vanish inside a swing in navigational demand, in either direction. Report them separately.

Next, check what else happened. A press hit, a launch, a seasonal swing, or a drop in category demand can push branded search up or down without changing how often an AI answer names you.

Google now gives you a native way to isolate branded queries. Search Console’s branded queries filter, announced on November 20, 2025, separates branded from non-branded queries without regex.

It has three main limits:

1. The data only exists from the point Google started tracking it, so you cannot backfill a baseline, as John Mueller confirmed in a LinkedIn thread.
2. Google’s documentation says the filter is unavailable for sub-properties and low-impression sites.
3. It will not split navigational from evaluative demand for you.

How do you capture self-reported attribution?

Self-reported attribution is the share of new opportunities in which the buyer names a source in their own words. Add two required free-text questions to your highest-intent form:

1. Where did you first hear about us?
2. What did you check before booking this?

In my opinion, free text beats a dropdown because a dropdown only lists the channels you already know.

Each month, sort the answers into a short list of groups (AI assistant, podcast, community, peer, event, search, other), count each group, and divide by the opportunities with a completed answer. In my experience, the second question is where AI research surfaces, because buyers remember consulting an assistant even when they cannot name the tool.

How do you size unexplained direct share?

Unexplained direct share is the portion of new pipeline that arrives as direct or branded, with no campaign, referral, or recorded prior touch behind it. Divide those opportunities by total new opportunities. It is a residual: what your tracking cannot explain. It should shrink as self-reported attribution fills the gap.

Template: the dark funnel scorecard

Copy this template and fill it monthly. It replaces the traffic verdict in your report. It does not add a total. The METER fragment for each row is in brackets. Recall appears on two rows because it has two halves: self-reported attribution and unexplained direct share

Use three decision rules when you read it:

1. Mentions up, branded lift flat: you are being mentioned without being remembered. Check citations by URL, because the evidence often sits on pages that never reach the buyer.
2. Branded lift up, mentions flat: something else is working, such as a podcast, an event, or a community. Check self-reported attribution before crediting AI.
3. Unexplained direct share falling while self-reported attribution rises: your measurement is improving, not your marketing. Do not read it as a performance gain.

I haven’t found a credible public benchmark for these five metrics, so build your own baseline over two to three cycles before you set targets.

How do you know a dark funnel metric is telling the truth?

A proxy earns trust when it moves before the outcome, repeats across events, drops when you stop the activity you think is causing it, and sits on a stable denominator. I call this the LEAD check: Lag, Echo, Absence, Denominator. A metric that fails any of the four gets reported as context, not as evidence.

In the fragmentation article, I suggested pairing AI visibility with branded search four to seven days later. Treat that as a hypothesis your data has to confirm, not a law of nature. The Similarweb study followed individual users. A weekly brand total is a blunter instrument, and blunt instruments need testing.

Proxies fail in predictable ways. A podcast appearance, a funding announcement, or a seasonal spike can lift branded search with no AI involved. A rise in mentions can reflect a changed prompt set rather than a changed market. So you test before you move budgets.

What is the LEAD check?

The eight-week minimum and the three-event rule are my rules of thumb, not statistical standards.

How do you run the Lag test in a spreadsheet?

1. Pull 12 weeks of two weekly series: the proxy (for example, share of AI mentions) and the outcome (branded clicks or new opportunities). Weekly totals remove the weekday pattern.
2. Calculate the correlation four ways: same week in levels, one-week lag in levels, same week in week-over-week change, and one-week lag in week-over-week change. Changes matter because two trending series correlate even when they are unrelated.
3. Compare each result to the threshold. With 12 weekly pairs, a correlation has to reach about 0.58, positive or negative, to stand out from noise at the conventional 5% level.
4. Weigh the lagged, week-over-week result most. It is the only one of the four that tests lead, and running four tests means one will occasionally cross the threshold by luck.
5. If the lagged test fails, do not present the metric as a cause.

Decide which tests you will run before you look at the results. Every extra test you try raises the odds of a false positive.

What does the check look like on a real brand?

Here’s what it looks like for Ulta Beauty in the US, measuring AI presence in ChatGPT. This is an illustration of the method on one brand, not a benchmark for any other. Every series covers the same 11 complete weeks, July 12 to September 20, 2026.

In September 2026, Ulta ranked first among beauty brands on the AI leaderboard, ahead of Sephora and CeraVe. It has an editorial content program, so AI answers have something to cite, and it has promotional branded demand, so that can confuse the results. You can read three of the five metrics from outside the company.

Similarweb’s own panel study also covers this: after a ChatGPT recommendation of Ulta, 7.6% of users visited Ulta, and 4.6% visited Sephora, while after a recommendation of Sephora, 7.9% visited Sephora and 3.3% visited Ulta.

The denominator matters here. The terms “ulta” and “ulta beauty” are 69% of Ulta’s branded clicks in September, and both are navigational. Evaluative and event terms such as “ulta birthday gift,” “ulta 21 days of beauty,” and coupon queries each hold under 2%. A lift computed on the head term mostly measures people who already shop there. The leaderboard’s top beauty prompt that month, “ulta beauty vs sephora, which is better?”, is the kind of question that precedes an evaluative branded search, and the head-term count cannot see it.

Now the Lag test, run the way the steps above describe it: visibility change this week against click change next week, on 9 pairs, where a correlation of about 0.67 is needed to stand out from noise (the 5% level, two-tailed, with nine pairs).

On ChatGPT, visibility moves first and narrowly clears the threshold (+0.69 vs. about 0.67). Visibility fell in the week of August 10 (29% to 11%) and again in the weeks of September 14 and 21 (31% to 5%). Branded clicks on the head term followed, from about 350K to 263K in August and from about 345K to 269K in September.

Run through the LEAD check and the result is mixed:

1. Lag passes by a small margin (0.69 vs. the 0.67 threshold).
2. Echo falls short: there are two events, but the rule of thumb requires three.
3. Absence has not been run. To run it, you need to stop an activity you think is working in one region and see whether results drop there while another region carries on as normal. I can’t do that to Ulta from the outside.
4. Denominator is shaky: ChatGPT visibility also fell for brands in unrelated categories in the same weeks, so the dips may reflect a platform-wide change and not something specific to Ulta.

Another factor could explain the pattern: Ulta’s clicks jumped 31% the week of July 26, right when its visibility peaked, and its sale event (“ulta 21 days of beauty”) is a top branded term. A sale can lift AI mentions and branded search together, with neither causing the other.

Rule out the platform shift and the sale before crediting AI.

The result: AI visibility looks like a signal worth tracking, but the evidence is not strong enough to say it drives branded clicks. This is what dark funnel measurement usually looks like: partial evidence, not proof. Run the checks first, then ask for a budget.

What does a failed check tell you?

A failed Lag test does not prove AI answers fail to influence buyers. It means you should not present that metric as a cause based on that evidence.

The usual suspects:

1. Your tracked prompt set misses the questions your buyers ask.
2. Much of your branded demand comes from existing users.
3. The effect is too small to see in a quarter.

The test cannot separate those explanations, and it does not need to. It tells you what not to claim, which is more than most dashboards do.

When do on/off tests and marketing mix modeling make sense?

Absence is the expensive check. Rand Fishkin’s point about incrementality is that measuring it means shutting an advertising channel off, often for months, which is why almost nobody does it. Run it only for the proxies you plan to move the real budget on.

Marketing mix modeling is the aggregate alternative when on/off tests are impractical. It estimates channel contribution from totals instead of individual tracking, which suits brand building that feeds invisible demand. It needs enough history and spend variation to work, so treat it as a quarterly validation layer, not a weekly dashboard.

Run the cheapest checks first.

1. Denominator takes about an hour: confirm you measure against the same base every time.
2. Lag needs a spreadsheet and data you already have.
3. Echo takes time, because you need more events to build up.
4. Absence is the most expensive, so save it for the metrics you plan to move budget on.

Which dark funnel metrics should you prioritize for your business model?

Prioritize by buying cycle length and committee size. Long-cycle B2B teams should lead with self-reported attribution and branded search lift. High-volume consumer and ecommerce teams should lead with share of AI mentions and engagement quality. Agencies should standardize on the full scorecard, so every client sees the same five numbers.

 

One scope note before the details. The Similarweb study covers US desktop users in finance, travel, and beauty, tracked over seven days, and excludes anyone who visited the brand in the prior four weeks. It is a clean read on new acquisitions in three consumer categories, not a law for enterprise software with a nine-month cycle.

The mechanism, a visit that arrives later through search or direct, isn’t category-specific. The size of the effect is.

What should B2B and high-ACV teams measure first?

Start with self-reported attribution. It is free, it captures channels no tool sees, and it works at any volume. If 69% of software buyers in G2’s survey chose a different vendor because a chatbot recommended it, the only instrument that hears about that is the question you ask on your form.

Add branded search lift next, but read it as a directional pattern across several events. At low opportunity counts, a single event is noise. Self-reported attribution also has different limits from the other four: it infers nothing from aggregates, and its weakness is buyer recall.

What should ecommerce and consumer teams measure first?

Start with Share of AI mentions, and analyze engagement next to it. In the Similarweb study, AI-influenced visitors viewed 12.0 pages and stayed 11.8 minutes on average, against 6.5 pages and 5.6 minutes for standard visits. The pattern fits visitors who are mostly sure of their decision and are verifying it.

The study did not test that explanation, so treat it as a hypothesis. If it holds, judge your landing pages on whether they offer proof, pricing, and a next step, not on whether they explain what you do.

What should you set up in the first 30 days?

Four changes cover it, and the first two need no new tooling: a required free-text field, a branded search baseline, a frozen prompt set, and citation counts next to traffic in every audit. Run the LEAD check on whichever metric moves first, before anyone reallocates budget because of it.

1. Week 1: Add the two free-text questions to your highest-intent form and make them required. Turn on Search Console’s branded filter so the baseline starts accumulating.
2. Week 2: Freeze a prompt set and a category definition. Count your share of AI mentions once, and record which denominator you used. For a first read on your category, the free AI leaderboard shows its most visible brands and most-cited sources.
3. Week 3: Add citation count next to the traffic number in every content audit and budget review. Require any recommendation to cut a page or reduce spend to carry a visibility trend line beside it.
4. Week 4: Fill in the scorecard for the first time. Add a column that logs events (launches, press, podcasts, community posts), and start the weekly series you will need for the Lag test.

Disclosure: Similarweb sells the panel and the tracking behind several figures in this article. The self-reported field and the branded search baseline cost nothing, and within a quarter they will tell you whether any of this applies to you.

Stop reporting traffic, start reading the pattern

Traffic did not become wrong. It became late. By the time you record a visit, the decision it records has mostly been made somewhere you were not watching, and a report built on the visit rewards whatever sits closest to the door.

The fix is not a bigger tracking stack. It is a smaller set of honest dark funnel metrics: five proxies with their own denominators, a scorecard read as a pattern, and a LEAD check that keeps proxies from passing as proof.

When mentions, branded lift, and self-reported answers move together across several events, you have a defensible read on dark funnel performance. When they diverge, you have found the fragment to work on.

If you want to track mentions and citations across engines without building it yourself, Similarweb’s AI Brand Visibility suite covers Share of AI mentions and citation-level data. Similarweb AI Studio connects those datasets with branded search and traffic, so you can ask a question in plain language and get the analysis, report, or dashboard back.

In a dark funnel, you cannot see the journey. You can still measure whether it is working.

FAQ

What are dark funnel metrics?

Dark funnel metrics are proxy measures of buyer influence that happens where no click is recorded: share of AI mentions, citations by URL, branded search lift, self-reported attribution, and unexplained direct share. They infer influence from aggregate signals instead of counting visits, so each one needs its own denominator and a validation step before it informs a budget decision.

What metrics should I track instead of website traffic?

Track five: share of AI mentions, citations by URL, branded search lift, self-reported attribution, and unexplained direct share. Each has its own denominator, so read them side by side and never add them into one score. Keep traffic as a diagnostic for where people land. It stopped being a verdict because only 8.8% of AI-influenced visits in Similarweb’s study arrived as AI referrals.

Is branded search a reliable dark funnel metric?

Branded search is a reliable directional signal and an unreliable proof. It lifts after AI recommendations, but also after podcasts, press, launches, and seasonal swings, so it cannot say which one worked. Separate evaluative terms from navigational ones, use Search Console’s branded filter for a baseline, and confirm the relationship with the LEAD check. For large budget moves, add an on/off test.

How long after an AI mention does branded search rise?

It depends on your category and your data. Similarweb’s study tracked visits for seven days after a ChatGPT recommendation among US desktop users in finance, travel, and beauty, and 55.9% of AI-influenced visits arrived through search. Seven days is a sensible starting window, not a guarantee. Run the Lag test on your own data across several events before you build reporting around any number.

Does a rise in direct traffic mean dark social is growing?

Not by itself. Direct traffic mixes typed URLs, bookmarks, dark social shares, and AI-influenced return visits. In Similarweb’s study, 19.9% of AI-influenced visits arrived as direct, against 38.8% of standard visits. Treat a direct spike as a prompt to check self-reported attribution and branded search, not as proof of any single cause.

How do you measure dark funnel performance without attribution?

Measure influence upstream and outcomes downstream. Upstream, track how often AI answers mention and cite you. Downstream, track branded search lift after visibility events, and ask buyers to name their sources in free text on your highest-intent form. With 68.01% of US Google searches ending without a click (SparkToro, January to April 2026), validate every proxy with the LEAD check before it informs budget.