What SEO Teams Should Test as AI Search Changes

SEO investment should follow evidence of useful visits and business value as search changes. Crawlability, indexation, useful content, internal linking, and accessible pages still determine whether content can participate in search. AI-generated answers add another layer in which a brand can be summarized, cited, or represented without receiving a click. The practical response is to manage five separate stages—eligibility, inclusion, visibility, referral traffic, and conversion—and connect them to qualified demand rather than treating rankings or citations as the final outcome.

Separate Current Evidence From Forecasts

The useful planning question is what to change on your own site. For the broader question of whether AI will replace SEO, see our existing explainer.

Google still requires an indexed, snippet-eligible page before it can appear as a supporting link in AI Overviews or AI Mode. Its official eligibility documentation says ordinary SEO practices remain relevant and special AI files or markup are unnecessary. These are platform requirements, not a forecast of your traffic or return on investment.

Keep three categories separate: confirmed product requirements, observations from a defined study, and predictions to test. A citation can establish exposure; it cannot by itself establish that someone paid attention or bought anything.

How AI search changes the discovery journey

The conventional search journey is relatively easy to picture:

Query → ranked result → website visit → action

A synthesized journey may look different:

Question → generated answer assembled from multiple sources → optional source visit → later action

Google describes AI Overviews as summaries that help users understand a topic and AI Mode as an environment for deeper exploration, reasoning, and comparisons. These experiences may use query fan-out, in which the system issues several related searches to retrieve information that supports different parts of the original request, as explained in Google’s AI-feature documentation.

In plain language, a person might ask one question about choosing payroll software, while Google retrieves information related to implementation, integrations, compliance, pricing, and company size. A strong page could surface because it addresses one of those supporting concerns, even if it does not repeat the user’s exact wording.

That does not justify creating a separate thin page for every conceivable variation. Google recommends distinctive, people-first content and warns against scaled pages created mainly to manipulate rankings or generative responses in its guide to optimization for generative AI Search features. Query fan-out can create additional retrieval paths; it does not establish that multiplying pages improves selection.

A better planning framework separates five stages:

Stage	Question	Observable signal	Limitation
Eligibility	Can the content be considered?	Indexed, crawlable, snippet-eligible page	Does not guarantee selection
Inclusion	Was the content used or linked?	Observed supporting link or citation	Observation may vary by query and user context
Visibility	Was the brand or page exposed?	AI impression, mention, or observed citation	Exposure does not establish attention or recall
Referral traffic	Did someone visit?	Referred session and on-site engagement	Referrer data may be incomplete
Conversion	Did visibility contribute to value?	Qualified action, opportunity, pipeline, or revenue	Attribution may be partial or multi-touch

Success at one stage does not guarantee the next. An eligible page may never be selected. A citation may receive no click. A visit may not convert. Conversely, an accurate no-click answer could influence a later branded search or direct visit that is difficult to attribute.

This is why the future of SEO is best managed as a visibility system. Teams need to understand where performance breaks down instead of using a single metric to represent the entire journey.

Check Eligibility Before Changing Content

For a small set of commercially important URLs, inspect crawl access, rendered text, the selected canonical, indexing and snippet eligibility. Check whether a shared template or access rule explains a problem across several pages. Publishing a page does not establish that Google indexed it.

Google’s SEO Starter Guide explains discovery, resource access and URL Inspection. Follow the existing technical checklist before buying an AI-specific optimization service. A technically eligible page can still be omitted from an answer; neither structured data nor an AI file guarantees selection.

The content advantage shifts from coverage to evidence

When a system can generate a competent generic summary, merely “covering the keyword” becomes less defensible. The stronger opportunity is to provide facts, experience, context, and decision support that a generic synthesis cannot create independently.

Prioritize material such as:

• Original research with a transparent method
• First-hand operational knowledge
• Expert analysis with named contributors
• Case studies that distinguish observations from conclusions
• Product or process comparisons using explicit criteria
• Implementation guidance grounded in actual constraints
• Clear limitations and cases where the advice does not apply
• Properly attributed facts and primary documentation
• Relevant screenshots, diagrams, demonstrations, or video

Presentation still matters. A direct answer under a descriptive heading helps a reader find the point quickly. Self-contained definitions reduce ambiguity. Comparison tables expose tradeoffs. Evidence placed beside the claim it supports makes verification easier. Descriptive captions and alt text provide context for useful media. Visible publication or revision dates can help readers assess whether time-sensitive information is current.

These are practical ways to improve clarity for people and systems, not a validated formula for earning citations on every platform. Formatting cannot compensate for weak evidence or guarantee selection.

Consider the difference between two software-category resources.

Generic version:

“The best project-management software helps teams collaborate, organize tasks, and improve productivity. Popular features include dashboards, automation, and reporting.”

That summary is easy to reproduce and does little to help a serious buyer decide.

Decision-support version:

• How implementation changes for a 20-person team versus a multi-department rollout
• Which identity, data, and workflow integrations matter
• What security documentation buyers should request
• How pricing changes with seats, usage, add-ons, and support
• Where migration typically becomes difficult
• Tradeoffs between configurability and administrative overhead
• Which buyer profiles are—and are not—a good fit
• Evidence from product documentation, expert interviews, and actual implementations

The second resource addresses uncertainty. It gives end users, technical evaluators, procurement teams, and financial decision-makers material they can use.

For B2B companies, one subject should often be mapped across those stakeholders rather than reduced to a broad, high-traffic article. An end user may need workflow examples; IT may need architecture and integration details; procurement may need contract and support information; finance may need pricing logic and implementation costs. These concerns can live in one well-designed resource or a tightly connected set of substantive pages.

AI can help organize research, identify gaps, draft outlines, or transform approved material into alternate formats. It should not remove editorial control. AI-assisted work still needs expert factual review, source attribution, originality checks, and brand-voice review. Consolidate commodity or duplicative pages rather than mass-producing lightly differentiated versions for every phrasing.

Fewer clicks do not mean no commercial value—but they change the risk

Current evidence supports concern about reduced click-through behavior, but not a universal prediction that AI answers will eliminate organic traffic.

In a March 2025 observational panel of 900 US adults, Pew Research Center found that users clicked a traditional result on 8% of visits where an AI summary appeared, compared with 15% of visits without one. A source cited inside the AI summary received a click on 1% of visits containing a summary. The results pages were reconstructed later using the same queries, and the study was not a randomized experiment, so it cannot prove that AI summaries caused the difference or forecast losses for a particular site. Those qualifications are part of the Pew analysis of AI summaries and subsequent browsing.

SparkToro separately reported a 68.01% zero-click estimate for Google searches in a Similarweb US desktop and mobile web panel covering January through April 2026. That is a panel estimate rather than a census or a probability that can be applied to an individual website. The SparkToro analysis of Similarweb clickstream data also cautions against direct historical comparisons involving different providers and populations.

The Pew and SparkToro percentages should not be subtracted, combined, or presented as a trend line. They involve different periods, populations, devices, data providers, and definitions. Neither source proves that AI Mode caused all observed zero-click behavior. Searches have long ended without external visits for reasons that include direct answers, navigational behavior, maps, calculators, knowledge features, and satisfied or abandoned queries.

Teams should plan for multiple journeys without trying to predict which will dominate:

Discovery scenario	Likely observable signals	Appropriate business response
Conventional click-through	Ranking, impression, click, session, conversion	Improve result appeal, landing experience, internal paths, and conversion
Low-click synthesized answer	AI impression, supporting link, mention, branded-search movement	Improve factual clarity, evidence, brand representation, and later-stage capture
Agent-mediated research	Variable mentions, referrals, direct or assisted actions	Maintain accurate decision data and test attribution across touchpoints

Effects will also differ by query. A simple definition may be satisfied on the results page, while a complex purchase, local service, technical evaluation, or high-risk decision may still require deeper research. Review performance by page type, intent, audience, device, and conversion path rather than multiplying all organic traffic by a market-wide zero-click estimate.

The new SEO scorecard: measure each stage separately

Google now provides dedicated reporting for organic impressions from AI Overviews and AI Mode. Its August 31, 2026 rollout update says the reporting is available to websites worldwide, with Search dimensions including page, country, device, and date. These impressions remain included in overall Web performance totals, as explained in Google’s generative-AI performance reporting announcement.

This is a meaningful improvement, but the boundary matters: it is visibility and impression reporting. It is not a complete AI-only dashboard for clicks, click-through rate, queries, conversions, or revenue.

Google documents several operational caveats. Dates use Pacific Time, the newest data may be preliminary, and the report may not appear when a property has insufficient impressions. Property-level chart totals and page-level rows use different aggregation, so their figures may not sum identically. Exports also convert displayed unavailable values to zero; analytics pipelines should preserve an availability flag rather than automatically interpreting every exported zero as a measured absence of impressions. These details are specified in Google’s generative-AI report definitions.

A useful scorecard should mirror the five-stage journey:

Stage	Core indicators	Measurement status
Eligibility	Crawl tests, index coverage, canonical state, snippet eligibility, rendering checks	Directly measurable
Inclusion	Observed links, citations, correct page selection	Experimental outside documented platform data
Visibility	Google AI impressions, brand mentions, citation accuracy and context; branded search as a demand signal	Mixed: direct for Google impressions, observational elsewhere
Referral traffic	Visits, engaged sessions, landing paths, returning visitors	Direct but potentially incomplete
Conversion	Qualified actions, assisted conversions, opportunities, pipeline, revenue	Direct or modeled depending on attribution

Citation monitoring should assess quality as well as quantity:

• Is the brand name correct?
• Is the description factually accurate?
• Is the appropriate page cited?
• Is the surrounding context favorable, neutral, or misleading?
• Does the answer distinguish the company from similarly named entities?
• Does the citation appear for a commercially relevant task?

Cross-platform citation frequency and sentiment remain experimental metrics.

Use a controlled prompt set instead of random screenshots. Record the platform, model or experience where visible, date, location, account conditions, prompt, response, citation, and factual accuracy. Treat changes as observations, not causal proof.

For Google’s AI-impression data, segment by page, country, device, and date. Compare movements with rankings, overall Web performance, branded searches, referral traffic, and conversions. Correlation can identify where to investigate, but it cannot establish that an AI impression caused a later sale.

Choose One Additional Channel to Test

AEO, GEO and AI-search optimization are industry labels, not one standardized system. Google’s guidance explains Google Search; it does not establish how other assistants or platforms select sources.

Keep Google’s search foundations in place, then choose another environment only when customer interviews, sales conversations or referral data show that buyers use it. Define one useful task and the evidence you can collect before assigning a budget. Monitor source accuracy and qualified visits; treat a few screenshots or mentions as observations, not proof of commercial impact.

A practical 90-day plan for adapting SEO

A 90-day program should establish the foundation, improve the most valuable content, and create disciplined experiments. It should not promise immediate rankings or AI citations.

Days 1–30: establish eligibility and baselines

• Select priority pages by commercial importance, not traffic alone.
• Audit crawlability, indexation, canonicalization, snippet eligibility, internal links, rendering, and accessible content.
• Review page experience and the usefulness of images or video.
• Verify business, location, service, and product information across relevant owned profiles and feeds.
• Record current rankings, impressions, organic visits, conversions, and available Google AI impressions. Include human attention and useful task completion, with bot filtering and missing-data coverage documented.
• Document analytics gaps, including unavailable or thresholded values.
• Identify pages that are duplicative, outdated, unsupported, or disconnected from conversion paths.

Days 31–60: strengthen evidence and decision support

• Add concise answers beneath descriptive headings.
• Bring in named expert input and first-hand operational knowledge.
• Attribute externally verifiable claims.
• Add comparisons, limitations, suitability criteria, and implementation details.
• Improve media where a diagram, screenshot, demonstration, or video reduces uncertainty. Test an interactive comparison or calculator only when a reader can use it to resolve a real decision; longer time alone can also signal confusion.
• Map content to end users, technical evaluators, procurement, and financial stakeholders.
• Add relevant next steps for readers at different buying stages.
• Consolidate commodity pages rather than multiplying keyword variants.
• Apply expert review, originality checks, and brand approval to AI-assisted work.

Days 61–90: test visibility and business impact

• Segment Google AI impressions by page, country, device, and date.
• Monitor a controlled set of representative prompts on one or two commercially relevant assistants.
• Record citation accuracy and context, not only mention counts.
• Inspect AI referrals, branded searches, direct visits, returning users, and assisted conversions.
• Connect qualified actions to opportunities and pipeline where attribution permits.
• Compare content cohorts over time without treating correlation as proof of causation.
• Define which tests to continue, change, or stop. Compare attention alongside task completion and qualified actions; do not reward a slower or more confusing page simply because time increased.

Use a clear action hierarchy:

Do now	Test cautiously	Avoid
Technical SEO and indexing checks	Cross-platform mention monitoring	Citation guarantees
Distinctive, evidence-rich content	Controlled prompt tracking	Mass-produced query variations
Accurate entity and product information	Assisted-demand analysis	Treating crawler visits as influence
Clear conversion paths	Incrementality or cohort analysis	Treating impressions as revenue
Measurement baselines and governance	Selected channel expansion	Universal optimization without audience evidence

Review the program quarterly. Interfaces, terminology, reporting, and assistant behavior are changing too quickly for a static annual playbook. The quarterly review should evaluate technical health, content quality, visibility, citation accuracy, traffic, qualified demand, platform priorities, and governance risks.

Teams can run this plan internally or evaluate a managed service. Request concrete deliverables, source-review responsibilities, publishing permissions and measurement definitions. Keep control over factual approval and assess the service against the same outcomes as an internal team.

At the end of the test, decide which changes helped visitors complete a useful task and which produced qualified demand. Keep those changes, revise inconclusive experiments, and stop work whose only result is a larger count of pages or mentions.