How AI Search Is Changing Visibility and What to Measure

If your average positions look steady while organic growth feels weaker, you may be measuring a journey that no longer happens in the same number of steps. A person can express a fuller need in one query, receive a synthesized answer, and skip follow-up searches that once gave you several chances to earn a click.

That changes visibility in two ways. Search sessions are becoming more compressed, and AI recommendations are less stable than conventional rankings. Your response should be an intent-based system that measures repeated presence, gives machines unambiguous evidence, and still helps a person make the decision in front of them.

Search demand can persist while the journey loses steps

Datos/SparkToro behavioral data from millions of users found that desktop Google searches per U.S. user fell by nearly 20% year over year. The decline in the EU and U.K. was much smaller, at roughly 2% to 3%. This is a per-user change, not proof that Google suddenly lost its audience.

The surrounding numbers make that distinction important. Traditional search remained about 10% of U.S. desktop activity through 2025. Dedicated AI tools accounted for only 0.77%, while Google AI Mode represented about 0.06% of U.S. desktop events by December. AI adoption is growing, but those shares are too small to support a simple story in which everyone abandoned Google for a chatbot.

These figures do not prove that AI caused every missing search. They are consistent with a more practical mechanism: AI answers and instant results can resolve part of a need before a person performs a second, third, or fourth query. Search remains central, but each session may generate fewer opportunities for publishers.

Query shape is changing at the same time. Six-to-nine-word searches are increasing rapidly in the U.S. Very long queries of 15 words or more remain uncommon and volatile, but they show that people are experimenting with more complete descriptions of what they need. You should therefore plan around the decision contained in a query, not just the keyword string that introduces it.

1. Choose one commercially meaningful decision. Examples include selecting a product for a constrained use case, deciding whether a service fits a particular situation, or comparing two approaches.
2. List the modifiers that change the answer. Audience, budget, compatibility, location, urgency, skill level, risk tolerance, and intended use can turn superficially similar prompts into different decisions.
3. Write down the facts required to answer each version. Include suitability, exclusions, specifications, limitations, evidence, availability, and the next action.
4. Map every important fact to a crawlable location. A claim should have a clear home on a page, not exist only in an image, sales call, private document, or advertising campaign.
5. Consolidate wording variants, but split genuinely different intents. If ten phrasings lead to the same criteria and answer, one strong resource can serve them. If the criteria change, create a distinct section or page rather than forcing every audience into generic copy.

This exercise gives you an intent map rather than another keyword list. It also exposes a common visibility gap: the page may mention the right topic while failing to provide the specific facts a search engine or AI system needs to answer the actual decision.

Measure AI visibility as repeated presence, not a fixed rank

An AI recommendation is generated for a particular request and context. It is not a stored, universally ordered result. Across nearly 3,000 executions of 12 identical prompts by more than 600 volunteers, an identical recommendation list appeared fewer than once in 100 responses. Getting the same list in the same order was rarer still, at fewer than once in 1,000.

A single screenshot therefore cannot tell you that your brand ranks third in AI search. It tells you that your brand appeared third in one response. Running the same prompt once more and reporting the better result is no more defensible; it replaces one anecdote with another.

The more useful signal is visibility percentage: how often your brand appears across a defined set of valid responses. Presence proved more stable than exact order, even when the lists themselves changed. Smaller niche categories tended to produce more consistent answers than large markets, so you should not compare percentages across unrelated categories as though they shared the same competitive conditions.

1. Define the prompt universe before collecting results. Select the audience, decision, market, language, and meaningful constraints. Do not add favorable prompts after seeing the outcome.
2. Create wording variants that preserve intent. Natural prompts can differ substantially in phrasing while expressing the same underlying need. Keep these in one family.
3. Separate prompts when the purpose changes. A general product recommendation and a recommendation for gaming, accessibility, enterprise security, or noise cancellation are different intent families if their selection criteria differ.
4. Repeat tests under documented conditions. Record the product or model, interface, date, locale, login or personalization state when known, exact prompt, and complete response.
5. Classify the outcome before calculating a rate. A passing mention, a direct recommendation, a citation, and an accurate description are not interchangeable forms of visibility.
6. Aggregate by intent family. Calculate repeated presence within each decision context before combining anything into an overall number.

There is not yet a validated universal minimum number of runs, and API output may not reproduce what a person sees in a consumer interface. Treat a small sample as directional. Keep the protocol consistent, retain the underlying responses, and widen the sample before making an expensive content or positioning decision.

You can still record list order for diagnosis. A persistent pattern may lead you to inspect what distinguishes frequently preferred brands. But exact position should not become the executive KPI, agency guarantee, or performance bonus when the output is inherently variable.

Make every important claim retrievable, specific, and verifiable

The next visibility problem is eligibility: can a system identify your entity, retrieve the relevant facts, and determine whether your offer fits the user’s constraints? A page can be persuasive to a person while remaining ambiguous to a machine because the product name changes between sections, limitations are missing, specifications live in images, or structured data conflicts with visible copy.

Moving from discovery to transaction inside one AI conversation is still a forecast rather than established behavior at scale. It is nevertheless sensible to make product and service information machine-readable now. The same cleanup also helps conventional search, feeds, internal search, accessibility, and human comparison.

Use this content pattern for each important decision page:

• Entity: State the exact product, service, organization, person, or location being described. Use the same canonical naming across headings, copy, metadata, and structured data.
• Direct answer: Address the central decision early. Say who or what the option is for, rather than making the reader assemble an answer from feature copy.
• Qualifiers: State compatibility requirements, exclusions, prerequisites, geographic limits, and material tradeoffs. Missing limits invite incorrect assumptions.
• Comparable facts: Present specifications, capabilities, availability, and policies in labeled text or tables where a comparison genuinely helps.
• Evidence: Add original measurements, first-party data, expert explanation, examples, or a documented method. Include enough context for someone to judge what the evidence does and does not establish.
• Freshness: Show when time-sensitive facts were reviewed, and correct outdated pages instead of allowing contradictory versions to coexist.
• Structured data: Apply the most specific relevant schema types and properties, using the same facts shown to the reader. Markup labels evidence; it does not replace evidence or make an unsupported claim true.

Generic summaries are easy to reproduce and hard to distinguish. Proprietary data and distinctive first-party content give other sites and AI systems information they cannot obtain from another lightly rewritten overview. The useful part is not merely owning data. You need to publish the method, scope, date, definitions, and limitations that make the result interpretable.

Specificity also protects brand accuracy. When your trial policy, service boundary, compatibility, or availability is unclear, a generative system may fill the gap with a category-level pattern that applies to competitors but not to you. Put the correction on the canonical page, align related pages and schema, and make the wording explicit enough to quote without reconstruction.

Do not create a separate thin page for every prompt variation. Build around meaning. A strong resource can answer several phrasings when the intended decision is the same, while modular sections can address the qualifiers that materially change the answer.

Treat video as visual, audio, text, and metadata

Video can supply evidence that prose struggles to carry: a product in use, a software workflow, a physical dimension, an expert’s explanation, or the exact state of an interface. AI systems can process visual frames, speech, on-screen text, and relationships between them. Some handle these streams together; others depend on separate recognition and transcription components. Either way, clarity determines how much useful information survives.

Optimize all four layers rather than uploading a polished file and relying on its title:

• Visual layer: Publish crisp 1080p video where practical. OCR can struggle with footage below 360p, and enhancement cannot reliably restore text that was never captured clearly. Use high contrast, bold readable type, and close enough framing for labels and interface states to be legible.
• Temporal layer: Keep a key object, label, or action on screen long enough to appear in sampled frames. Rapid cuts may look energetic to a person while causing an automated system to miss the one frame that establishes the fact.
• Audio layer: Use clear speech, identify speakers, reduce competing noise, and align narration with the action on screen. Deliberate pauses can separate important statements and reduce ambiguity.
• Text layer: Provide human-verified captions and a transcript. A transcript gives text-dependent systems access to the substance and reduces errors introduced by automatic speech recognition.
• Metadata layer: Use accurate titles and descriptions, then add applicable VideoObject markup. Properties such as hasPart, transcript, and interactionStatistic should describe real, visible content and verified data.

Review the finished video without sound, then review only the audio and transcript. If either version loses the core claim, the layers are not reinforcing one another. Fix the asset itself before adding schema; metadata cannot rescue an unreadable demonstration, an incorrect caption, or a missing limitation.

Use a scorecard that separates exposure, accuracy, and value

Traffic remains useful, but it no longer describes the whole journey. An answer can mention your brand without linking to it, cite you without recommending you, recommend you inaccurately, or send a visitor who converts. Those are different outcomes and should occupy different rows in your reporting.

Key takeaways

• Fewer searches per person do not mean Google has become irrelevant; they mean each journey may contain fewer opportunities.
• An AI list position is an observation from one response, not a durable rank.
• Measure repeated brand presence across defined intent families and documented conditions.
• Separate mentions, recommendations, citations, accuracy, and business outcomes.
• Improve visibility eligibility with explicit facts, distinctive evidence, consistent structured data, and machine-readable media.

A practical scorecard can use the following definitions. Set the inclusion rules before testing, and keep the denominator visible beside every percentage.

Store the prompt and complete response behind every AI observation. Also retain the model or product, interface, collection date, locale, and personalization state when known. Compare like with like. If a platform changes, preserve the old series and label a new baseline instead of hiding the discontinuity inside a blended average.

Do not force no-click visibility into a revenue number you cannot defend. Report correlation as correlation, keep attributable conversions separate, and use brand visibility trends to decide where to investigate. The purpose of the scorecard is to improve decisions, not manufacture certainty from a probabilistic system.

On your next reporting cycle, start with one high-value customer decision. Build its prompt family, collect a documented baseline, identify the most obvious evidence or accuracy gap, and correct that gap on the canonical page. Then rerun the same protocol. That gives you a repeatable visibility practice while the interfaces, models, and search journeys continue to change.

References

• Search Engine Land – SEO in 2026: Expert Predictions for a Transforming Landscape
• Search Engine Land – Study Reveals AI Recommendations Rarely Repeat: What It Means
• Search Engine Land – U.S. Google Searches Drop: The Impact of AI on User Behavior
• Search Engine Land – Boost Your Brand: Optimize Videos for AI Search