Part six
Measurement: what you can actually measure, and where the instruments mislead you
Six platforms use the word "impression" to mean six different physical events. Your analytics has a channel for AI traffic and will not tell you what it counts. And the single best-evidenced link between a marketing number and revenue is about ratings.
-
13 May
2026, the day GA4 gained an AI assistant channel. There is no AI assistant data before it
Google 1
-
1,000
Daily consenting users your site needs before Google will model the visitors who refused. A restaurant will never reach it
Google 2
-
0.17%
Of an average site's traffic came from AI assistants, across 3,000 sites, February 2025
Ahrefs 3
-
5 to 9%
Revenue increase per one star on a review site, from tax records, and only for independents
Luca, Harvard 4
-
6 of 6
Platforms whose definition of an impression differs from every other platform's
Each platform's own docs 5
-
Google will not publish what counts as AI traffic
The source list it links to contains search, social, shopping and video. There is no AI category in it. 1
-
No AI company documents its referrer behaviour
All three publish detailed crawler docs. None mentions what it sends when somebody clicks a citation. 6
-
A zero in your report can mean "too small to show"
Google's thresholds are "system defined. You can't adjust them," and it publishes no number. 2
-
Instagram deleted impressions entirely
Deprecated on 21 April 2025. Its reach figure is annotated by Meta as an estimate. 5
- Google, GA4 release note 13 May 2026 and default channel group documentation. 1st
- Google, GA4 documentation on modelling and thresholding. 1st
- Ahrefs, 3,000 sites, February 2025. 18 months old. 2nd
- Michael Luca, Harvard Business School, on Yelp ratings and restaurant revenue. 1st
- YouTube, Meta, X, LinkedIn and Google Business Profile documentation. 1st
- OpenAI, Perplexity and Anthropic documentation, checked 8 August 2026. 1st
Your analytics has an AI channel and will not tell you what is in it.
This is the question every owner is asking right now: is anything coming from ChatGPT? Here is exactly what the instrument does, and exactly where it stops.
What Google added, and when
On
13 May 2026
Google added an AI assistant channel to
Analytics. The mechanism matters: it is not a referrer test. Google rewrites the traffic medium to
ai-assistant
"if the referrer matches a list of AI Assistants," and the channel then matches
on that medium.
7
Two consequences follow immediately. There is no AI assistant data before 13 May 2026 , and Google does not state that the rule was applied to historical data. And because the classification happens upstream, on a list, you are not looking at referrers. You are looking at Google's opinion of which referrers are AI assistants.
Nobody on the other side documents it either
We checked OpenAI's, Perplexity's and Anthropic's own documentation. All three publish detailed crawler documentation. None of them mentions what referrer, if any, is sent when a person clicks a citation in an answer. The hostname list that circulates in every guide to this subject is third-party observation of undocumented behaviour, and it can change without notice. 8
How much traffic is this, actually
The best-sampled measurement we found: 0.17 percent of an average site's traffic , across 3,000 sites, February 2025. That is 18 months old and it is a vendor's own customer base. 9
Two claims circulate about the quality of that traffic, that AI visitors are 4.4 times or 2.5 times more valuable. Neither publishes a sample size, a measurement window or a denominator, and the first is drawn from marketing and SEO topics, which is not a restaurant. We are not going to repeat either as a figure. 10
- Google, GA4 release note, 13 May 2026, and default channel group documentation. Analytics Help pages carry no dates. 1st
- OpenAI, Perplexity and Anthropic crawler and bot documentation, checked 8 August 2026. 1st
- Ahrefs, 3,000 sites, February 2025. 2nd
- Semrush and Similarweb. Neither publishes a sample size or a denominator for the claim. 2nd
Why a small site sees zeros where a big one sees numbers.
Three mechanisms in Google Analytics work against small sites specifically, and all three are documented by Google. None of them is a fault in your site.
-
01
Thresholding, not sampling
Sampling only begins at 10 million events in a query, which a restaurant will never see. Thresholding is the one that bites: Google withholds rows when the counts are small enough to identify individuals. Its own wording is that the thresholds are "system defined. You can't adjust them," and it publishes no number for them. A zero in your report can mean "too small to show," not "none." 11
-
02
The modelling floor
When visitors decline cookies, Google can model the missing data. To do it, your site needs "at least 1,000 daily users sending events with analytics_storage='granted' for at least 7 of the previous 28 days." A single venue will not reach that, ever. So for a small business, consent refusals are missing data, not modelled data. 12
-
03
Basic consent mode
Most small-business cookie banners are configured in basic mode, which blocks the tags entirely until consent. That disables modelling altogether, regardless of the floor above. 12
What this means when you read a report
Take a month where your Analytics shows no traffic at all from a channel you know produced enquiries. Three explanations are available before you reach "it did not work": the row was thresholded, the visitors declined cookies and were not modelled, or the referrer was stripped before it arrived, which is the subject of the whole dark social chapter earlier in this document.
- Google, GA4 documentation on data sampling and data thresholding. Analytics Help pages carry no publication dates. Retrieved 8 August 2026. 1st
- Google, GA4 documentation on behavioural modelling and consent mode. 1st
- Google, GA4 custom channel group documentation, and the release note of 13 May 2026. 1st
The four free instruments, and where each one stops.
Every one of these is documented by the company that operates the surface it measures, which makes them worth more than any third-party dashboard. Each has a boundary worth knowing before you rely on it.
| Instrument | What it actually gives you | Where it stops |
|---|---|---|
| Google Search Console | Impressions, clicks, click-through rate and average position for your pages in Google Search, including a separate generative AI performance report showing impressions in AI Overviews and AI Mode | The AI report is impressions only, and the two features cannot be separated. AI Overviews and AI Mode traffic is folded into overall web numbers with no way to filter it out. And an AI Overview impression only counts if the link is "scrolled or expanded into view," so a collapsed citation registers nothing. Google's own API documentation states it "does not guarantee to return all data rows but rather top ones" 14 |
| Bing Webmaster Tools | An AI Performance report, public preview since 10 February 2026, showing "when your site is cited in AI-generated answers across Microsoft Copilot" | Microsoft surfaces only. The grounding queries it shows are explicitly described as "a sample," with no published rate. It is still the only place any company tells you when its AI cited you 15 |
| Google Business Profile | How people found you, what they did, calls, direction requests and website clicks from your listing | Retention is limited and Google documents a change to how direction requests are counted without giving a date for it . For most venues this is the single highest-value free instrument in this list, because it measures the surface where local decisions actually happen 16 |
| Google Postmaster Tools | Your email spam complaint rate, domain and IP reputation graded Bad, Low, Medium or High, authentication results and delivery errors | Spam rate counts only mail that reached an inbox, so a sender being auto-filtered can show a healthy complaint rate while barely arriving. Watch reputation, not spam rate 17 |
- Google, Search Console Help and Search Console API documentation, the latter dated 20 May 2026. Help pages carry no dates. 1st
- Microsoft, Bing Webmaster Blog, 10 February 2026. 1st
- Google, Business Profile performance documentation. Retention period and the date of the directions counting change could not be established. 1st
- Google, Postmaster Tools documentation. No publication date on the page. 1st
Six platforms use the word "impression" for six different events.
Every one of these definitions is the platform's own. Put them side by side and the practice of adding up reach across platforms, or comparing one platform's engagement rate to another's, stops making sense.
| Platform | What it counts, in the platform's own words | What that does to a comparison |
|---|---|---|
| YouTube | A thumbnail shown with at least 50 percent visible for over one second, with a documented list of placements that are excluded | The strictest definition here. YouTube also states that its own click-through rate cannot be derived from views divided by impressions, which is exactly the sum most people do 18 |
| Meta, Facebook | Content that "entered a person's screen." No duration requirement at all | A Facebook impression and a YouTube impression are not the same event and are not within an order of magnitude of each other in effort 19 |
| Impressions were deprecated entirely on 21 April 2025. The primary metric is now views. Meta annotates Instagram reach as estimated | Any Instagram figure from before April 2025 is measuring something the platform no longer reports. And its reach number is a modelled estimate by Meta's own label 19 | |
| X | Impressions are explicitly not unique. The same user viewing a post twice counts as two | The only platform of the six that says this outright. It makes X impressions structurally larger than a unique-user metric 20 |
| "Number of impressions." The documentation defines nothing further | Undefined by the platform. Whatever you conclude from a LinkedIn impression, LinkedIn has not told you what it is 21 | |
| Google Business Profile | Deduplicated to one per unique user per day | The most conservative counting rule in the table, and the one attached to the surface where local decisions actually happen. A GBP impression is worth far more than a Facebook one and the number will always look far smaller 22 |
- YouTube Help, impressions and click-through rate documentation. 1st
- Meta, Insights and metric documentation, including the Instagram impressions deprecation of 21 April 2025. 1st
- X Help Centre, post analytics definitions. 1st
- LinkedIn analytics documentation. 1st
- Google Business Profile performance documentation. 1st
- TikTok's metric definitions could not be retrieved at any stable public URL on 8 August 2026, so TikTok is absent from this table rather than guessed at. n/a
One study links a marketing number to restaurant revenue. Here it is.
We went looking for published research connecting any marketing metric to revenue for a small or local business. We found one good paper. That is not a literature, and the honest thing is to say so.
What the paper did
Michael Luca, then at Harvard Business School, matched restaurant review ratings against Washington State tax records , using a regression discontinuity design across 3,582 restaurants . Revenue was measured from the tax filings, not from a survey or a platform's own reporting. 24
The finding: a one-star increase in rating produced a 5 to 9 percent increase in revenue . And the qualifier that matters more than the headline: "this effect is driven by independent restaurants; ratings do not affect restaurants with chain affiliation." 24
Two limits, stated plainly. The data ends in 2009, which makes it very old. And it is about rating . It says nothing about reach, followers, impressions or engagement.
And what we did not find
- No other study linking a marketing metric to revenue for a small or local business, with a stated method
- Nothing at all on the accuracy or bias of asking customers how they heard about you. Two separate research passes came back empty. It is recommended everywhere and validated nowhere
- Nothing on the practical measurement of word of mouth for a local business
- Michael Luca, Reviews, Reputation, and Revenue: The Case of Yelp.com , Harvard Business School. Regression discontinuity on Washington State tax records, 3,582 restaurants, data to 2009. 1st, peer reviewed
- Cloudflare blog, July 2025, retrieved 8 August 2026. Searched for the ratios it is cited for and found none. 1st
Six claims checked against the source.
Measurement is the part of marketing that sounds most like accounting and behaves least like it.
| The claim | Verdict | What the source actually says |
|---|---|---|
| You can see how much traffic AI is sending you | A number, not an audit | The channel exists, added 13 May 2026, and it classifies traffic against a list Google does not publish. The source categories spreadsheet Google links contains search, social, shopping and video, with no AI category. No AI company documents its referrer behaviour either. 26 |
| Reach can be added up across platforms | Six different events | YouTube requires a thumbnail 50 percent visible for over a second. Meta requires only that content "entered a person's screen." X counts the same person twice. LinkedIn defines nothing. Google Business Profile deduplicates to one per person per day. Instagram deleted impressions altogether in April 2025. 27 |
| A zero in analytics means nothing happened | Often means too small | Google withholds small rows and states the thresholds are "system defined. You can't adjust them," publishing no number for them. A small site is far more likely to hit thresholding than sampling. 28 |
| Consent mode models the visitors who refuse | Not at your size | Modelling needs "at least 1,000 daily users sending events with analytics_storage='granted' for at least 7 of the previous 28 days." A single venue will not reach it, and basic consent mode disables modelling entirely. 28 |
| Search Console keeps 16 months and shows 1,000 rows | Not in the documentation | Neither figure appears in Google's own Performance report documentation. The API limit is documented, at 1 to 25,000 rows with a default of 1,000, and Google states the API "does not guarantee to return all data rows but rather top ones." 29 |
| Reviews drive revenue | Measured, once, on tax records | A one-star increase produced a 5 to 9 percent revenue increase across 3,582 restaurants, measured from Washington State tax filings, and "this effect is driven by independent restaurants; ratings do not affect restaurants with chain affiliation." The data ends in 2009 and nobody has repeated it. 30 |
- Google, GA4 release note 13 May 2026 and channel documentation; OpenAI, Perplexity and Anthropic documentation. 1st
- Each platform's own metric documentation. 1st
- Google, GA4 thresholding, sampling, modelling and consent mode documentation. 1st
- Google, Search Console Help and API documentation, the latter dated 20 May 2026. 1st
- Michael Luca, Reviews, Reputation, and Revenue: The Case of Yelp.com , Harvard Business School. 1st, peer reviewed