With AI, awareness and consideration move inside the chats: the user arrives at the service already oriented, often in the conversion phase, after a filtered version of the information the site would have provided first-hand.
The data is consistent. Those arriving from AI referral convert more than organic, although estimates vary widely and are almost all from vendors: Semrush (June 2025) speaks of a rate 4.4 times higher; Adobe measures retail conversions from AI referral 54% higher than non-AI (May 2026 data, was +42% in the first quarter); Ahrefs reports that 0.5% of sessions from AI generate 12.1% of registrations — a differential of about 23×. To be treated as orders of magnitude, not measurements: AI traffic remains under 2% of visits (Datos, first quarter 2026) and attribution is broken upstream.
The estimates are plausibly conservative: many receive the recommendation in chat, search the brand on Google, and buy — conversions attributed to branded organic, with the AI contribution invisible to standard analytics. The Similarweb clickstream study quantifies it: those exposed to a ChatGPT recommendation have 2.5 times more probability of visiting that site within seven days, and 55.9% of those visits come from branded search (versus 40.4% of standard traffic). These visitors also arrive more decided: 12.0 pages in 11.8 minutes, versus 6.5 pages in 5.6 minutes for ordinary visitors (The Downstream Impact of AI Visibility, June 2026 — US desktop only, three verticals, declared correlation and not causality).
The top of funnel becomes less accessible: the AI Overview appears on about half of queries (Google declares ~50%, BrightEdge measures 48-50% in February 2026). When it appears, organic CTR drops — by how much is the subject of twelve independent analyses that agree on the direction and diverge on the magnitude: from −15% (Amsive, 700,000 keywords) to −89% (DMG Media, evidence filed with the UK CMA), with the methodologically most solid measures around −35/−65% (Ahrefs on 300,000 keywords: −34.5% in April 2025, −58% in December 2025 in position 1; Seer Interactive on 25.1 million impressions: from 1.76% to 0.61%, i.e. −65%). The effect goes deep: position 2 −50.8%, position 3 −46.4%, and even position 10 loses 19.4%. In AI Mode over 90% of sessions close without a click (93%, Semrush). On the editorial side the aggregate effect is heavy: across over 2,500 monitored outlets, Google referrals fall 33% globally and 38% in the US (Reuters Institute/Chartbeat, 2026).
The same shortening happens in the inbox. Apple Intelligence summarizes the message before opening, effectively replacing the preheader; Gemini in Gmail summarizes after, and to do so opens the email — making the open rate even less reliable than it already was. A new figure emerges, the “informed non-opener”: the user gets the value of the message from the summary and never opens it. It is the same mechanism as the AI Overview applied to a channel thought to be safe — an intermediary inserts itself between content and recipient, produces a synthetic version, and measurable interaction disappears.
A useful exception to know: on e-commerce and shopping queries the AI Overview appears on only 3.2% of searches, after starting from 29% — Google pulled back because generative answers did not convert into sales. The phenomenon now touches navigational and brand queries, much less transactional product ones. Inside the site remain conversion and retention, where attention should be focused; loyalty and advocacy keep traditional dynamics.
Effect on the professions
Digital Strategist / Growth Marketer — Organic acquisition enters structural crisis: work on how to be cited by AI engines and hold the post-search moments. Being cited has a measurable effect: on queries with AI Overview, cited brands get about double the organic clicks per impression (Seer Interactive, 2026). Paid follows the same trajectory: Google introduces ads in AI Mode and AI Overviews, Perplexity sells sponsorships in answers — with open questions on formats, measurability, and the user’s ability to distinguish organic and sponsored recommendations. → Useful resource: How AI Is Changing Search Behaviors — NN/g
Digital / Web Analyst — Attribution models break, but not in the way usually said. The measured fact is that AI-influenced traffic arrives much less as “direct” (19.9%, versus 38.8% of ordinary traffic) and much more as branded search (55.9%, versus 40.4%): the user receives the recommendation in chat and then searches the brand on Google (Similarweb, June 2026). The effect is real — 2.5 times more probability of a visit within seven days — but invisible to a last-click model, which attributes it to branded organic. You need dedicated channel groups for AI referrers, awareness that “direct” and branded organic are contaminated containers, and an upstream measure of visibility in AI answers. 2026 news: that measure finally has an official source. Since 3 June Search Console offers the Search Generative AI performance reports, with impressions inside AI Overviews, AI Mode, and the generative features of Discover — for now only impressions, without clicks, CTR, average position, or queries, and with the three surfaces aggregated. → Useful resources: Your analytics are lying — Similarweb via PPC Land; Search Generative AI performance reports — Google Search Central
AI Visibility Analyst — An autonomous measurement discipline is born, with its own object: whether and how much the brand appears in generative answers. The research published by SparkToro and Gumshoe on 27 January 2026 — the first to verify whether these tools are consistent enough to produce valid metrics — says precisely what can and cannot be measured. The design: 600 volunteers ran 12 prompts on ChatGPT, Claude, and Google (AI Overview and AI Mode), for 2,961 total executions, repeating each prompt 60-100 times; the analysis resumes a Carnegie Mellon methodology on LLM consistency.
What is not measurable: position. Repeating the same question a hundred times, each answer differs on three dimensions — which brands appear, in what order, and how many (sometimes two or three, sometimes more than ten). The probability that ChatGPT or Google return the same list twice is less than 1 in 100; that they return it in the same order, less than 1 in 1,000. It is not a defect: they are probabilistic engines, designed to generate a different answer every time. A clear consequence follows — any tool that sells a “ranking position inside AI” is selling an artifact. And a risk to know, because the mechanism is manipulable: if an answer is not liked, just re-run the prompt until the desired one comes out. It is the same dynamic as the unscrupulous SEO vendors of twenty years ago, and it applies to vendors as much as to internal reports.
What is measurable: share of presence. Across dozens or hundreds of prompts run many times, the frequency with which a brand appears is stable and informative. In the study, an agency appeared in 85 of 95 Google AI answers; a hospital in 69 of 71 ChatGPT answers (97% visibility), while being the first mention in only 25 cases. The percentage of presence says how much an entity is inside the model’s consideration set — which is the useful information — while the order says nothing.
How to calibrate the number. Expected visibility depends less on brand strength than on the breadth of the competitive space: where candidates are few (cloud providers for SaaS, dealers in a single city) the top names reach 90-100%; in broad spaces (brand design agencies, newly released novels) the best stop at 30-40%. Comparing your number with that of another sector means nothing: the benchmark is internal to the space.
The prompt set. People do not write similar prompts even with the same intent — average semantic similarity 0.081 across 142 formulations of the same need (see Intent-based interaction). Since the model captures intent beneath the phrasing, synthetic prompts turn out to be an acceptable proxy for real ones: the set must be built by intent, broad (dozens or hundreds of prompts), versioned, and kept stable over time, otherwise the measured variations are your own and not the model’s. It must be replicated across multiple platforms, given the de-concentration of the market.
Operating rules. Repeat each prompt at least 60-100 times before considering a data point valid. Reason on averages and long windows, never on weekly readings. Do not produce or accept rankings. Verify outputs against other sources: in the study the models repeatedly recommended entities that no longer exist — inactive influencers, closed accounts, defunct companies.
Two volatilities not to confuse. The one described so far is synchronic: the same question, at the same moment, produces different answers, and is governed with repetitions. There is a second, diachronic: according to Profound’s longitudinal analysis of 240 million citations, 40-60% of cited domains change within a month and 70-90% comparing two semesters. This is governed with time series. A report that does not distinguish the two is summing noise and signal.
Cautions on the source. One of the authors works in a sector company (Gumshoe), a conflict openly declared in the piece; the study is not peer-reviewed and the survey is from November-December 2025. The authors themselves leave the decisive questions open: how many executions are needed for statistical significance, whether API calls reproduce the variety of real users, how many prompts are needed per sector.
What to ask a vendor. The AI tracking market is already worth over $100 million a year in estimated spend and is consolidating: Profound raised $96M at a $1 billion valuation (February 2026), the category went from about 7 to over 150 products in ten months, and in June 2026 Adobe acquired Semrush. Before signing, four questions: how many times do you run each prompt; how do you build and update the set; do you publish the methodology in verifiable form; do you sell ranking positions — in which case the answer has already arrived. Search Console has offered since 3 June the first official source, but limited to Google surfaces and impressions only: it integrates monitoring, does not replace it, and watch out for lock-in on proprietary metrics not comparable between vendors.
The boundary with other roles. This role measures the output (what the model says); Social Listening guards the input (the sources and conversations that feed that answer); Digital PR acts on those sources; the Web Analyst measures downstream traffic. Four links of the same chain: a loss of share of presence is almost always explained by looking upstream. → Useful resources: AIs are highly inconsistent when recommending brands or products — SparkToro; LLM monitoring tools — Semrush
CRM / Email & Lifecycle — If a growing share of messages is read in summary, the message must be designed so the summary works: essential information and call to action in the first lines, one message per email, clean semantic structure (real headings, no text inside images), a subject that does not depend on the preheader to be understood. On the measurement side, the open rate must be demoted from KPI to weak signal: move judgment to clicks, conversions, and replies, and consider a declining open rate not necessarily a worsening. To test explicitly: how your email appears once summarized by the main clients. → Useful resources: AI summaries in email clients — Stripo; AI-Mediated Inbox Environments — PMA
Product / UX Designer and Content Designer — The landing page must presuppose an already oriented user, not an exploratory one: they arrive with fewer questions and more expectations, and go twice as deep if the content holds up.