Search traffic fell 34 per cent. The measurement finally caught up
For two years the honest answer to how do we measure AI search was a shrug and a spreadsheet. That answer expired on Monday.
The number that ends the argument#
A fox does not argue with a river. It trots upstream along the bracken until it finds the crossing. Publishers have spent two years arguing with a river, and the crossing has finally appeared, which makes ai visibility a measurable discipline rather than a nervous hunch.
First the river. On 31 July 2026 Jim VandeHei published Chartbeat data showing that Google search traffic to publishers fell 34 per cent over the past year. His framing was blunter than the number: Google has basically stopped sending people to websites for answers and information, and is instead answering them on its own platform, in its own words.
The two-year figures are the ones to take to a board meeting, because they show the loss is not evenly spread. Small publishers lost 60 per cent of search referrals over two years. Medium publishers lost 47 per cent. Large publishers lost 22 per cent.
Track that gradient carefully, because it is the whole ai visibility strategy in one shape. A large publisher with a known name held on to more than three quarters of its referrals. A small one lost nearly two thirds. Whatever is deciding who survives is closer to recognition than to relevance, and that is a brand problem wearing a technical costume. The finding was widely picked up, including in Slashdot's coverage the same day.
Why ai visibility replaces the click as the metric#
If the answer arrives before the click does, counting clicks measures the leftovers rather than the quarry. What matters is whether the model quoted you while forming the answer, and until this week that was extremely difficult to see without paying a specialist vendor for it.
Free tooling just ate a paid category#
On 3 August 2026 Microsoft Clarity split its AI citation reporting into branded and non-branded queries. Search Engine Land's report on the release lists four additions: branded labels on individual queries, a Share of Authority breakdown by query type, dashboard filters to compare visibility by category, and citation analysis that separates brand strength from topic discovery.
That distinction is the entire ai visibility game. A citation on a query containing your brand name proves the model knows who you are. A citation on a generic query proves the model thinks you are the best available source on a subject. The first is reputation. The second is acquisition, and conflating them has been making ai visibility reports useless for two years.
One number, no decision in it
We appeared in 240 AI citations this month, up 18 per cent. Nobody can tell whether that is growth or simply people looking us up by name.
Two numbers, one obvious action
Branded citations 190, non-branded 50. The model knows us but rarely picks us on open questions, so the work is topical depth, not brand awareness.
The pace of release tells its own story. Per the Microsoft Clarity blog, Citations reached general availability on 13 May 2026, Topic Insights arrived on 9 July, citation queries by topic on 22 July, and the branded split on 3 August. The gaps between releases are collapsing.
For anyone who has been quoted a five figure annual fee for llm visibility monitoring, this is the moment to renegotiate. A free analytics product now reports branded and non-branded citation splits, topic grouping and competitor citations. Specialist vendors will need a sharper answer than dashboards to justify the invoice.
What to actually put on the report#
Tooling does not make a metric. Plenty of teams will now export an ai visibility count, put it on slide four, and be no wiser. A useful ai visibility report answers three questions, and it needs both Clarity and Search Console to answer them.
Start with the diagnostic everyone skips. Open the Search Console performance report and read impressions and clicks together. If impressions hold while clicks fall, you are being summarised rather than demoted, which is a content problem. If impressions fall too, you have a ranking problem and no amount of citation work will fix it.
| Impressions | Clicks | What it means and what to do |
|---|---|---|
| Holding or rising | Falling | You are being summarised. Work on quotability, self-contained sentences and entity clarity. |
| Falling | Falling | A ranking or indexing problem. Fix eligibility first, because citation work cannot rescue a page nobody retrieves. |
| Rising | Rising | Nothing is wrong. Resist the urge to change anything on the strength of a headline. |
| Falling | Holding | Your remaining audience is loyal and probably branded. Protect it and build topical depth separately. |
Then add the off-site layer, which most reports still ignore entirely. Google made Search Console platform properties available worldwide in late July, covering Instagram, TikTok, X and YouTube accounts, per Search Engine Journal and Google's own platform properties guidance. Query-level data for content that never touched your domain is a genuine gap closed.
Getting cited is dull work, and it works#
Measurement without a lever is just anxiety with a chart, and ai visibility deserves better. The levers that move ai search visibility have not changed much, and they remain as unglamorous as a moonlit prowl along a familiar hedgerow.
Open every page with a short, direct answer to the query it targets, in plain text, with the term used naturally and no marketing preamble in front of it.
A summary lifts a sentence, not a section. Each claim must make sense with nothing above or below it, which is the practical reading of Google's helpful content guidance.
A model cannot safely repeat a superlative but it can safely repeat a number with a named source attached at the point of the claim.
Ship Article, FAQPage and Organization markup so the system knows who is speaking and can attribute the answer to a verifiable organisation.
Give each subheading a fragment identifier so a citation can point at the paragraph rather than the page, and so your analytics can see which passage earned it.
Google's own documentation supports each of those. Its guidance on AI features in Search and its helpful content guidance both point the same way, and the structured data documentation covers the entity layer.
That is why folkfox keeps putting brand strategy next to SEO and GEO services rather than in a separate proposal. The technical work makes you eligible. Recognition decides whether you are chosen.
One piece of housekeeping from the same week#
A smaller change deserves a mention in any ai visibility audit, because it removes a box people have ticked since 2007. On 31 July 2026 PPC Land reported that Google dropped the documented requirement to block internal site search result pages, a rule that dated to the original webmaster guidelines.
John Mueller still recommends blocking them, for two reasons that have nothing to do with policy: infinite crawl space and server load, and search boxes being spammed with casino, pharma or adult content. So the technical recommendation stands even though the rule has gone, which is a distinction worth making to any client who reads the headline and asks why you have not changed anything.
Volatility watchers had a busy week too, with tools tracking ranking movement between 1 and 3 August. Nothing was confirmed by Google, and the appropriate response to unconfirmed volatility is to hold changes and document, not to chase. Search Engine Land maintains a useful running record of how AI Overviews behave across these episodes.
Non-branded citation share
Citations on queries that do not contain your brand, as a share of all citations. This is the acquisition number.
Impression to click ratio
Track the direction, not the level. Impressions holding while clicks fall means you are being summarised.
Anchored passages shipped
Count subheadings with fragment identifiers, because a citation that can point at a paragraph is worth more than one that cannot.
Off-domain queries seen
Queries surfacing your social and video content, now visible through Search Console platform properties.
Those four ai visibility lines start at zero deliberately: they are a template rather than a benchmark, and the first month's job is simply to populate them honestly. Any agency handing you a filled-in version of this in week one is showing you a projection, not a measurement.
The pattern beneath all of it is the one folkfox described in the shift from clicks to visibility, and it plays out sector by sector, from FinTech and Web3 brands going invisible in AI search to telehealth providers winning citations. Ai visibility measurement arrived late. It has arrived.
Frequently asked questions#
What does AI visibility actually mean?
It means how often AI answers and assistants cite your content when forming a response, rather than how many clicks they send you. Because a summary can answer a query without a click, citation frequency is a better signal of whether your content is influencing the answer than session counts are.
How much has AI search reduced publisher traffic?
Chartbeat data reported at the end of July 2026 showed Google search traffic to publishers falling 34 per cent over a year. Over two years the losses were uneven: 60 per cent for small publishers, 47 per cent for medium and 22 per cent for large ones.
Axios synthesises Pew, Chartbeat, Similarweb and CMA evidence on the same dataset, bolstering the number that ends the argument.
Why does the branded and non-branded split matter so much?
Because they measure different things. A citation on a query containing your brand shows the model knows you exist. A citation on a generic query shows it considers you the best source on a topic. Only the second one represents new demand, so reporting them as a single figure hides the decision.
Do I still need a paid AI visibility tracking tool?
Increasingly less so for basic reporting. Microsoft Clarity now offers branded and non-branded citation splits, topic grouping and competitor citations without charge. Paid tools need to justify themselves on coverage across more assistants, historical depth or workflow rather than on dashboards alone.
How do I tell whether I am being summarised or demoted?
Read impressions and clicks together in Search Console. If impressions hold steady while clicks fall, your pages are still being retrieved and the answer is being given without a click, which is a content problem. If impressions fall as well, that is a ranking or eligibility problem and needs fixing first.
Should I still block my internal site search pages from Google?
Yes, in practice. Google removed the documented requirement at the end of July 2026, but its own advice still recommends blocking them because they create infinite crawl space and because search boxes get spammed with casino, pharma and adult content that can become a site quality issue.
Read more on this topic#
SEO vs GEO: why your best pages miss AI citations entirely
The structural reason strong pages get skipped, and what to change on them.
Read the pieceEU Fintech & Web3 brands in AI search
What the same measurement problem looks like in a regulated category.
Read the pieceHow EU Telehealth Providers Win AI Citations
A worked example of the dull, effective moves that earn a citation.
Read the pieceThe regulator stopped reading your ad and started reading your funnel
Why quotable, self-contained compliance sentences now serve two audiences at once.
Read the piece
Want the reporting built rather than described?
folkfox builds ai visibility reporting for brands in regulated categories: honest baselines, branded and non-branded splits, and content changes that actually move the citation count.