Why visibility and visits need separate measures in AI search

Two developments verified in September

First, Ahrefs' original France study ↗ compared Search Console data from the 28 days before AI Overviews launched in July 2026 with the nine days after. The analysis retained 963 domains. In the group highly exposed to AI Overviews, median click-through rate fell from 2.29% to 1.76%, a relative decline of 23.1%. Ahrefs reports a 2.3% increase over the same period for its low-exposure control group. The 23.1% figure is the observed change, not a control-adjusted estimate. A short French panel cannot be copied into a forecast for every sector or country.

Second, Google Search Central's September 24 announcement ↗ added a filter for image-based web searches in Search Console. It covers Lens, Circle to Search on Android, image uploads, and Chrome's image search. The filter is available in the Search results performance report and the report for Generative AI features. Rollout is gradual worldwide, and data appears when a property receives traffic from those queries. This creates another observation window; it does not measure brand preference or sales contribution.

The two events do not prove the same outcome. One is a short before-and-after observation in a specific market; the other is a reporting feature for a new search path. Together they raise a practical question: why treat the possibility of being found and the act of visiting as one number?

Identify the decision before changing the campaign

A decline in visits is not, by itself, a reason to rewrite the campaign message. Separate three possibilities first.

Consider a fictional luggage brand whose campaign promise is “freedom to leave lightly.” The ad need not carry every specification, capacity figure, and return condition. Search answers and the product page should connect that promise to real use cases, dimensions, storage, delivery, and returns. Record separately whether the bag appears in image search, which traveler an answer says it fits, what a visitor compares on the site, and how people remember the brand. This is a measurement-design example, not a real campaign result.

Three strategic options for the next quarter

1. Prioritize recovery of qualified visits. Choose this when a customer needs to do something on the site—check a price, stock, booking, or tool—and CTR declines repeatedly for comparable query intent. Also check whether the destination page lets the customer complete the promised action. 2. Prioritize answer quality. If people compare options before clicking and AI explanations misstate when the brand fits, improve product information, examples, and evidence of differentiation. The target is not a citation count by itself, but an accurate reason for the right person to choose. 3. Hold the budget decision and repair measurement. If countries, devices, and query groups are mixed or AI exposure cannot be separated, fix the baseline first. Do not use the average from a short French study as a target for a Korean account.

These are alternatives, not a checklist to deploy everywhere at once. Pick one based on whether the campaign aims to create visits, awareness, or inquiries and where its evidence chain breaks.

A four-week validation design

In week one, fix the customer questions and pages, as well as country, language, device, and search type. Establish separate baselines for standard search, image-based search, and any available AI-feature reporting. Record metric definitions and whether each report is actually available. If a value is absent, label it “unmeasured” or “relevant traffic not confirmed,” rather than entering zero.

In week two, improve the explanation on landing pages for one question group only. Keep the campaign's emotional message while adding the use case, attributes, and conditions that support it. Retain a comparison question group or unchanged pages, and check whether their exposure conditions are comparable.

In weeks three and four, compare impressions, CTR, and post-visit actions separately. If the objective is an inquiry, cart, or purchase, keep its event definition fixed. If understanding or preference is the objective, design a separate short survey or interview. Do not attribute a change to AI when product prices, promotions, seasonality, or search features changed at the same time. Without repeated observations or a sound comparison, leave the result as a diagnostic hypothesis.

Success is not a single score. A more accurate recommendation with fewer clicks can mean different things under different objectives. Likewise, more clicks are not a success if the destination contradicts the promise made earlier.

Limits of application

The Ahrefs figures come from a provider-run study of 963 French domains, a particular exposure group, and a short period. We did not independently reanalyze its raw data and cannot generalize it to Korean clients. Google's filter is rolling out and may show data only for eligible traffic; it does not mean every account already has comparable observations. We did not measure a client campaign, brand perception, or revenue for this report. The four-week plan is an untested method, not an established performance prescription.