Google’s Search announcements at I/O 2026 changed more than the appearance of a results page. The company made Gemini 3.5 Flash the global default in AI Mode, introduced an AI-powered Search box that accepts text, images, files, videos and Chrome tabs, and made it easier to continue from an AI Overview into a conversation.

For publishers, the immediate temptation is to “optimise for AI” everywhere. That is too vague to manage and too risky to measure. The better response is a 90-day observation plan that separates what Google announced from what actually changes in your audience, pages and revenue.

Start with what is known

Google says the new Search box can expand for longer questions, suggest ways to formulate intent and accept multimodal inputs. It also says conversational follow-ups retain context and can surface increasingly relevant supporting links.

The following implications are reasonable inferences, not guarantees:

  • some journeys may move from short keywords to longer, task-shaped questions;
  • a page may be discovered later in a conversation rather than on the first query;
  • images, video and files may become part of the question, not just the answer;
  • a useful source may earn fewer visits from simple facts but more qualified visits from complex decisions.

Do not turn those possibilities into invented certainty about ranking factors or citation formulas.

Capture a baseline before changing content

Export at least the previous 90 days for organic search. Preserve weekly data by page and query rather than only site totals.

Record:

Layer Minimum baseline
Search visibility Impressions, clicks, CTR and average position
Landing pages Organic sessions, engaged sessions and key events
Business outcome Affiliate clicks, leads, subscriptions or sales
Content state Title, main heading, update date and material page changes
Device and market Device, country and language where volume supports analysis

Annotate releases, redesigns, tracking changes and major content updates. Otherwise a template change can be mistaken for an AI Search effect.

Group queries by job, not by word count

“Long-tail” is not a sufficient category. Build query groups around what the searcher is trying to complete:

  • definition: what is this feature?
  • selection: which tool fits my budget or team?
  • comparison: how do two options differ under a real constraint?
  • procedure: how do I complete a task safely?
  • verification: is a claim, price or source current?
  • troubleshooting: why did a workflow fail?

Conversational Search may change the phrasing while preserving the job. A query cluster lets you observe that continuity.

For a software review site, compare “best AI video tool” with task-shaped variants such as “tool for four eight-minute avatar videos under $50.” The second query exposes the evidence the page needs: credit arithmetic, duration, resolution and revision risk.

Measure pages through a four-question review

Review each priority page weekly or monthly, depending on traffic.

1. Is demand changing?

Use impressions by query cluster. Falling clicks with stable impressions may indicate a result-page change, weaker snippet appeal or a more complete answer on Google. Falling impressions may indicate demand, indexing, competition or relevance changes. Do not diagnose both with the same remedy.

2. Are visits becoming more qualified?

Compare clicks with outcomes. Calculate:

qualified outcome rate = desired organic outcomes / organic landing sessions

If sessions fall 15% but newsletter subscriptions remain stable, the business effect differs from a 15% fall in both. Zero-click behaviour matters, but a publisher should not optimise only for visit volume when the commercial goal is qualified action.

Use a small worksheet rather than relying on the headline traffic number:

Fill in Current 90 days Previous 90 days
Organic landing sessions
Desired outcomes
Qualified outcome rate

A falling session count beside a stable or improving outcome rate requires a different response from a decline in both columns. Check seasonality and tracking changes before attributing either result to AI Search.

3. Does the page provide extractable evidence?

Inspect whether an editor—or a search system—can identify:

  • the direct answer to the page’s question;
  • current facts with primary sources and checked dates;
  • calculations with visible assumptions;
  • a named limitation;
  • who is accountable for the page;
  • a distinction between tested findings, vendor claims and editorial inference.

This is not a request to write robotic answer blocks. It is a quality-control test for ambiguity.

4. Does the click deliver more than the summary?

A page becomes expendable when its entire value fits in a generated paragraph. Add material that earns the visit: a downloadable checklist, a worked cost model, original screenshots, a reproducible test, a decision table or an update log.

The addition must be real. Expanding 800 generic words into 1,800 generic words does not create information gain.

Build an AI Search observation sheet

Use one row per material observation:

Date Query cluster Page Change observed Evidence Competing explanations Action Review date

Hypothetical example—the dates below illustrate how a team could schedule an observation and follow-up; they are not BenPicks performance data:

18 Aug 2026 | HeyGen plan selection | /ai-video/heygen-creator-vs-pro-youtube/ | impressions stable, CTR down, affiliate clicks stable | Search Console and analytics | title change on 12 Aug; stronger result feature visible | no rewrite; monitor two weeks | 1 Sep 2026

The “competing explanations” field prevents a familiar SEO mistake: assigning every movement to the newest industry story.

Update pages with controlled changes

When evidence supports an update, change one meaningful layer at a time where practical:

  1. correct stale facts and sources;
  2. improve the opening answer and decision context;
  3. add missing evidence, calculations or first-hand material;
  4. make headings reflect actual reader jobs;
  5. verify structured data matches visible content;
  6. record the change and wait for enough data.

Do not add FAQ sections only to repeat the article. Do not manufacture quotations, product tests or author experience. Google states that using generative AI does not itself determine quality, while generating many pages without added value can violate scaled-content policies.

Treat multimodal search as an editorial requirement

If users can ask with images, video and files, publishers should make their own non-text material understandable and useful:

  • write descriptive alt text for informative images;
  • add captions that explain what a screenshot proves;
  • provide video transcripts and meaningful chapter labels;
  • place charts near the analysis they support;
  • expose units, dates and assumptions in tables;
  • avoid embedding the only important fact inside an unlabeled image.

This improves accessibility and human comprehension even if no special AI Search benefit follows.

What not to promise in a report

Avoid statements such as:

  • “AI Search traffic is fully isolated” when the available report does not support it;
  • “this page was cited because of schema” without direct evidence;
  • “mentions are replacing links” based on a few manual searches;
  • “longer prompts require longer articles” as a universal rule;
  • “we recovered traffic” before normal volatility and content changes were considered.

Manual spot checks can document how results look, but they are not stable rank tracking. Personalisation, location, language and continuing conversation context can change the output.

A 90-day operating rhythm

Weeks 1–2: preserve the baseline, validate analytics events and choose ten commercially important pages.

Weeks 3–6: classify query jobs, record visible changes and repair only clear accuracy or value gaps.

Weeks 7–10: improve two or three pages with measurable additions—such as a worked calculator or documented comparison—and annotate the release.

Weeks 11–13: compare query-cluster visibility, qualified outcomes and revision results. Keep, reverse or extend changes based on evidence.

The goal is not to finish “AI SEO” in a quarter. It is to build a measurement system that survives the next interface change.

The practical conclusion

Google’s 2026 Search box makes questions richer and journeys more conversational. Publishers cannot control where a source appears in that journey, but they can control whether a page is current, attributable, measurable and worth opening after a summary.

Preserve the baseline. Track reader jobs. Add evidence a generated answer cannot honestly invent. That is a stronger strategy than rewriting the whole site around an unverified theory of AI citations.

Sources