Two things happened in the same week, and read together they change how a beauty brand should report AI visibility for the rest of the year. Google began rolling out AI Search performance reporting inside Search Console globally, so impressions across AI Overviews and AI Mode are now measurable natively for the first time. And independent tracking put the share of US ChatGPT answers carrying an actual web citation at under 7% for the year. One of those gives you real data. The other tells you that most of the AI visibility dashboards on the market are describing something their own methodology created.
We have been running the same measurement cleanup across client reporting all summer, from the scoreboard metrics in the rise of ACOS, MER, and weighted CAC down into the input layer. AI visibility is the newest number to arrive on that sheet, and it is arriving with the weakest definition of any metric we track. Worth fixing before it starts driving budget.
The 7% number, and why it breaks a whole product category
When someone types a question into ChatGPT, the model usually answers from training rather than going out to the live web. Independent tracking puts the share of US ChatGPT answers containing an actual web citation at under 7% this year. No retrieval happened, so there was nothing to cite, and no brand got credited or linked.
That matters because several platforms sold as AI search visibility trackers work by forcing search mode behind the scenes. They submit a prompt in a way that guarantees retrieval, then report whichever citations come back. The output looks like a legitimate visibility report. It describes a mode of the model that fewer than one in fourteen real answers uses. A team optimising against that report is optimising against an artifact of how the tool queried the model.
The two questions to ask your vendor this week
Ask whether the platform forces search mode to generate citation data, or samples how the model responds by default. Vendors that force search are not being dishonest, since forcing retrieval is often the only reliable way to get a citation to measure at all. The problem is a report that does not say so. If the methodology is not disclosed anywhere in the documentation, that absence is your answer.
Then ask whether the platform can separate two numbers: citation rate under forced search, and mention rate in the model default answer to the same prompt. These describe different problems. A brand that shows up well under forced search but never gets named in a default answer has a retrieval presence and no training presence. A brand invisible in both has neither. The fixes are not the same, and a single blended score hides which one you have.
What we found auditing our own visibility
Pennock is named first in the AI answer for "best beauty marketing agency," described as the pick for emerging to scaling beauty brands. Pennock is absent from the AI answer for "skincare marketing agency paid media" for the second consecutive month, where nine other agencies are named.
Same brand, same site, same month, two adjacent commercial queries, opposite results. A single visibility score would have averaged those into a number that told us nothing. The prompt level split is the finding, and it is the only view that points at work to do.
What Search Console now gives you, and what it does not
The Search Console rollout is the genuinely new measurement asset. AI performance reporting shows impressions across AI Overviews, AI Mode and other generative surfaces, broken out by page, country, device and date. This is first party Google data about Google surfaces, which makes it the most trustworthy AI visibility input any brand currently has.
It is also narrow. It covers Google, not ChatGPT, Perplexity or Claude, and it reports impressions rather than whether your brand was named inside the answer. Treat it as the reliable floor of your AI reporting and treat everything else as directional until the methodology is disclosed. A brand that reports Search Console AI impressions honestly and says nothing about ChatGPT is in better shape than one reporting a confident cross platform score it cannot source.
The publisher opt out is now a live decision
Google shipped a Search Console control that blocks a site from its generative AI features without affecting rankings in traditional Search. The trade is explicit: you lose AI impressions and the traffic behind them.
For most beauty brands this is an easy no. Discovery is moving toward AI surfaces, not away, and opting out to protect content from summarisation forfeits the exposure at the same time. The brands where it deserves a real conversation are those whose organic traffic is concentrated in long form editorial that AI Overviews answer completely, where the summary satisfies the reader and the click never happens. That is a per page question, and Search Console now gives you the page level data to answer it rather than guess.
Where AI visibility actually shows up in a beauty funnel
Consideration queries, not brand queries. Nobody asks an AI assistant to describe a brand they already buy. They ask which retinol is worth the money, which acne routine works on skin of color, which brand to try instead of the one that irritated them. Those are the prompts where a model either names you or names nine competitors, and they map closely to the mid funnel terms we already track in organic search.
Practically, that means the pages that earn AI mentions are the same ones that earn featured snippets: specific, structured, answerable, with a real position rather than a hedge. Our 2026 skincare advertising benchmarks get cited because they publish numbers with a stated source and date. Undifferentiated category pages do not get cited because there is nothing in them a model needs.
AI discovery is not replacing paid, it is repricing it
The reason to care about any of this is that paid acquisition costs are not falling while discovery shifts. If an AI answer becomes the first place a customer hears about a competitor, you are buying the same customer later in their consideration and paying more for the privilege. AI visibility is a top of funnel media cost you can influence without a bid.
Google also put numbers on the paid side of the same week. Per data reported across the trade press, human voice in a YouTube asset correlates with 12% higher conversions, brand shown within the first five seconds with 4% higher, and text overlays with 3% higher. Google cites Nielsen research attributing 49% of campaign ROI to creative. These are correlations, not tested lifts, so use them as a brief checklist rather than a forecast. The point stands either way: the controllable levers keep turning out to be creative and content, not media price. We made the same argument channel by channel in our comparison of Meta and TikTok for beauty acquisition, and the auction context sits in our Meta advertising cost benchmarks for beauty brands.
Marketplace discovery is moving on the same axis
Two more data points from the same week fit the pattern. TikTok Shop reached roughly 2% of total US online retail spending in July 2026 on Consumer Edge card transaction data, up from 1.2% a year earlier, which puts it ahead of Target, Costco and Home Depot individually. Q2 spend grew 60% year over year across every income tier, and shoppers over 35 are now the fastest growing demographic on the platform. Separately, OpenAI pinned Shopping to the ChatGPT sidebar with dedicated landing and search pages.
Read those together and product discovery is fragmenting away from your own site faster than most beauty brands have budgeted for. The over 35 detail is the one that should change a plan, because TikTok Shop has been funded as a Gen Z channel while the fastest growing spenders are the cohort buying mid and premium skincare.
The AI visibility audit, in five steps
- Turn on the Search Console AI performance report and baseline AI impressions by page. This is your only first party number.
- Write down 10 consideration prompts a real customer would type, not 10 keywords. Include the ones where a competitor should win.
- Run each prompt yourself in the default mode, with no forced search, and record whether your brand is named. Manual, unglamorous, accurate.
- Ask your visibility vendor the two methodology questions above, in writing, and put the answer in the report footer.
- Report the trend line, not the score. A consistent methodology tracked over three months survives scrutiny. A single number does not.
Four ways teams misread AI visibility
First, treating a vendor score as a market share figure, when it is a sample of prompts the vendor chose. Second, reporting an isolated number to a client or a board, which invites a follow up question no tool can answer with precision. Third, optimising for citations when the model is not retrieving, which puts effort into a mode most answers never enter. Fourth, and most expensive, cutting content investment because AI Overviews reduced clicks on a handful of informational pages, when those same pages are what earn the mentions on consideration prompts. Each error makes the reported number look tidier while making the underlying position worse.
One row to add to your reporting sheet this month
Add four cells beside the organic numbers you already report: Search Console AI impressions, count of tracked consideration prompts where the brand is named in a default answer, count where a competitor is named and you are not, and the vendor methodology in plain language. That last cell is the one that keeps the other three honest.
AI visibility is worth measuring. It is not yet worth a confident single score, and any vendor selling you one owes you the methodology first. If you want the same efficiency discipline applied to the paid side, our CPM calculator and ROAS calculator handle the arithmetic, and our primer on what return on ad spend actually measures is the place to start if the scoreboard still needs cleaning up.