What Agencies Should Measure Now That Google Stopped Sending Clicks

Measure six layers, in this order. First, whether AI engines name and cite you: AI Mention Rate, AI Citation Rate, AI Share of Voice. Second, whether that visibility lifts demand you can see: Branded Search Lift and Google Business Profile actions. Third, whether the traffic that does arrive is good: AI Referral Quality and AI Referral Conversion Rate. Fourth, whether it becomes qualified leads and booked appointments. Fifth, whether those produce revenue at an efficient ROAS, CAC, and LTV:CAC. Sixth, underneath all of it, how much of each number is observed versus modelled. Average position and raw session counts are no longer the headline. They describe a channel that is quietly being answered before the click.

Last reviewed June 2026. The benchmark bands below are operating starting points, not guarantees; tighten them against 60 to 90 days of your own baseline. Every external claim links to its primary source.

For twenty years the agency business ran on one question: did Google send a click. Rankings produced clicks, clicks produced sessions, sessions produced conversions, and the monthly report was a tidy line from position to revenue. That chain is breaking at the first link, and a lot of agencies are still reporting as if it holds. We run measurement across ten AI platforms and roughly sixty thousand data points a day on a single source of truth, and the rebuild below is what that vantage point now demands.

Why did the click stop being the unit of value?

The click stopped being the unit of value sometime in the last two years, and the data is no longer ambiguous. Ahrefs' December 2025 analysis, published February 2026, found that the presence of an AI Overview cut the position-one organic click-through rate by about 58 percent for the queries where it appears. Pew Research, studying real browsing, found people click a traditional search link in just 8 percent of searches that show an AI summary, versus 15 percent without one. The answer now appears above the results, the user reads it, and the visit never happens. The brand was still surfaced. The recommendation was still made. None of it shows up as a click.

The reporting makes the gap worse. Google now folds AI Overviews and AI Mode traffic into the standard Web search type in Search Console rather than exposing a separate AI mention or citation metric. An AI Overview occupies a single position with all its links sharing it, and each AI Mode follow-up question is counted as a new query. So the visibility you create inside these answers is largely invisible to the dashboard you have always used to prove your work. The report says traffic is flat or down, while the real story is that the client is getting recommended more than ever in places the old tools cannot see.

Three forces are driving this, and each one degrades a different part of the report.

Zero-click search removed the click but not the exposure. Generative answers, featured snippets, and local packs increasingly satisfy the query on the results page. Google is explicit that there are no special requirements, and no need for files like llms.txt, to appear in its generative experiences: the lever is the same content quality and entity health it always was. Google also does not tell you to ignore rank, but it warns that average position is a complex metric that is easy to misread, and recommends watching how it moves over time rather than reading a single number. The dashboard keeps shrinking in smaller ways too. Google stopped showing FAQ rich results in May 2026 and is removing the FAQ report and its Search Console API support across June to August 2026.

Observability loss made the surviving numbers softer. Google Analytics moves on 15 June 2026 to Consent Mode as the single control for how your data is collected and used, taking that role from Google Signals. Where users decline analytics cookies, Google models their behaviour and models the key events they would have produced. A ROAS figure built half on modelled conversions is not the same asset as one built on observed purchases, and a report that does not distinguish the two is quietly overselling its own certainty.

Attribution divergence made the numbers argue with each other. GA4 answers acquisition at three different scopes: user acquisition for how new users first arrived, traffic acquisition for how sessions started, and event scope for how key events are credited. User and session scope use paid-and-organic last click; event scope uses the model you select, defaulting to data-driven attribution. Those scopes legitimately produce different answers for the same channel. Teams that never wrote down which one is the source of truth end up debating dashboards instead of diagnosing marketing.

What should I measure now? The short answer

Measure the chain from visibility to revenue, not the click in the middle of it. The table below is the whole answer in one view: fifteen metrics across six layers, each with what it answers, how it is calculated, which system owns the truth, and a starting benchmark band to tighten with your own data. The sections after it explain the layers that are new or most misread.

The discipline is to read this top to bottom, not to pick one number. A high AI Mention Rate with a flat Qualified Lead Rate is a positioning problem. A healthy Booked Appointment Rate with a collapsing Consented-traffic share is a measurement problem wearing a performance costume. The point of the stack is to tell those two apart.

How do I measure AI visibility?

Start measuring the thing the old dashboard cannot see. If clients are now recommended, cited, and compared inside AI answers, that exposure is the new top of the funnel, and it needs its own metrics. No platform hands them to you. Google folds AI traffic into Web reporting with no citation report; OpenAI's ChatGPT Search shows users inline citations but gives site owners no analytics; Perplexity documents crawler controls and a partner revenue program, not per-prompt visibility. The clearest proof that this is a build-it-yourself problem: vendors such as DataForSEO now sell an LLM Mentions API that has to define "citation" versus "mention" itself, because the platforms do not. So you build the dataset: a controlled library of prompts, run on a schedule across the AI surfaces, every answer stored and parsed.

Three metrics come out of that library, and the distance between the first two is the whole game.

AI Mention Rate is the share of monitored prompts where the brand is named at all. AI Citation Rate is the share where one of the client's own pages is actually linked as a source, which is a higher bar and a better predictor of durable visibility. AI Share of Voice is the client's slice of all mentions across a defined competitor set, judged against an equal-share baseline so "good" means beating the field, not beating last month. None come from GA4 or Search Console. They are the closest thing we have to a ranking report for the generative era.

Two bridge metrics connect that visibility to something the business can feel. Branded Search Lift tracks growth in branded query demand against a seasonal baseline, because the most common signature of strong AI exposure is not a click from the answer, it is a person who reads it and then searches the brand by name a day later. The Zero-Click Influence Score is a deliberately derived composite that blends the AI signals with branded search and local action into one index anchored at 100. It is not a vendor-native metric and we never present it as precision. It is an honest trend line on influence that produces no click, which is exactly the influence the old dashboard zeroes out.

Which mid-funnel metrics survived the click's decline?

Local action and lead quality carry more weight now, not less. When the top of the funnel stops producing clicks, the mid-funnel signals that do survive become disproportionately valuable, and the honest move is to lean on them harder.

Google Business Profile is the clearest example. Its Performance API reports daily counts for calls, website clicks, direction requests, bookings, and conversations, plus monthly search-keyword impressions. Those are real actions taken by real people who decided, often straight off a zero-click surface, to contact the business. For a local client, a rising action rate against stable impressions is frequently a truer read on momentum than any sessions chart.

Referral quality matters more than referral volume. Google reports that when people click from a results page showing an AI Overview, those clicks are higher quality, meaning users spend more time on the site. That matches what we see: AI-referred traffic stays relatively small but behaves better than the site average. An agency that judges AI referrals on raw volume will dismiss the best-converting traffic on the site. Judge it on engaged-session rate, contact-page reach, and conversion against the site benchmark, then track AI Referral Conversion Rate separately so quality and outcome never blur together.

How do I prove revenue, not traffic?

The bottom of the funnel is where the report should land, and it belongs to the CRM, not to GA4. Qualified Lead Rate, Booked Appointment Rate, and revenue with ROAS and CAC are owned by the system that knows what a lead is worth. The standard economic sanity check is still an LTV-to-CAC ratio near 3:1. GA4 supplies behavioural context and offline-event backfill; the CRM supplies the truth.

This matters even in the channels with the most explicit intent. WordStream's 2026 benchmarks put the average paid-search conversion rate at 8.18 percent and cost per lead near 67 dollars: proof that even bottom-funnel demand is something you pay more to convert each year, which is exactly why the report has to end in qualified pipeline and revenue rather than in clicks. The agencies that keep their clients through this transition are the ones whose monthly report opens with qualified leads and revenue and treats traffic as the supporting cast it has become.

How do I know which numbers to trust?

The discipline that ties the stack together is honesty about certainty. Because so much of the funnel is now modelled, derived, or observed through a controlled prompt set rather than logged directly, every number deserves a confidence tag sitting next to it. That is not a hedge. It is the most valuable thing an agency can add to a report in 2026.

In practice it means three diagnostics ride alongside the media metrics. Modelled-conversion share tells leadership how much of the conversion count is inferred rather than observed, which matters more after the 15 June 2026 consent change tightens what gets collected. Consented-traffic share tells them how much of the behavioural data is even eligible to be seen. CRM match-back rate tells them how much of the claimed revenue actually reconciles to closed business. When those diagnostics are visible, a softening number prompts the right question, which is whether performance moved or whether observability moved. When they are hidden, the agency is one consent-rate shift away from taking credit for a modelling artifact, or blame for one.

What this scorecard does not tell you

Honesty about limits is part of the authority. Three caveats keep this from being oversold.

The AI-visibility metrics are derived observability, not official platform metrics. Google, GA4, and the Business Profile API do not report AI Mention Rate, AI Citation Rate, or Share of Voice; they have to be built from a controlled prompt library with stable run conditions and human QA, and they are only as good as that library is representative. Treat them as a measured estimate, not a meter reading.

The benchmark bands are starting points, not law. They come from public platform documentation and vendor studies, not from a universal standard for your vertical, your city, or your service line. The right thresholds are the ones you set after 60 to 90 days of your own baseline, then revisit quarterly.

Attribution is still a judgment call. There is no single model that perfectly assigns credit across a zero-click answer, a branded search, a phone call, and a delayed close. The honest move is to report observed and modelled side by side, lock your attribution rules in writing, and validate with the CRM rather than trusting any one platform's number.

The scorecard that replaces rankings

Stop leading with average position. The click it used to predict is evaporating, and it tells the client nothing about whether the business grew. The scorecard that replaces it runs in a single chain: AI visibility, to branded demand, to local action, to qualified leads, to booked appointments, to revenue, with a confidence level on every tile. The top of that chain is built from a prompt library you own. The bottom is built from a CRM the client owns. The middle is the local and behavioural signal that survived the click's decline.

This is more work than pulling a rankings export, and that is the point. The measurement gap that opened over the last two years is the clearest line in our industry between agencies that are keeping up and agencies that are about to be surprised in a renewal meeting. The clients are already being recommended in places the old dashboard cannot see. The only question is whether their agency can prove it.

For the systems behind this stack, see how we track visibility across ten AI platforms, monitor sixty thousand data points a day, and build a single source of truth that joins all of it to revenue. For the failure mode this is built to prevent, read what happens when conversion tracking breaks.