What is agentic search?

It is search carried out by an AI agent that runs several steps for the user, such as searching, reading pages, comparing options and sometimes completing a task, instead of returning a list of links.

How will AI agents change off-page SEO?

The reasoned expectation is that agents shortlist a few options, so independent reviews, comparisons and mentions decide who is in the set. This has not yet been measured in any public study.

Do backlinks matter to AI agents?

Indirectly, as far as anyone can tell. Agents retrieve pages through search indexes, and links still help pages rank there. A link also sits inside a mention, which is what AI systems appear to weigh.

Is there data on how many people use AI agents to shop?

Not that we can verify. Figures circulate in vendor blogs without a traceable source, so we do not repeat them.

What should I do now to prepare for agentic search?

Keep business facts consistent everywhere, earn independent coverage and reviews, and make key pages light enough for a machine to read quickly. All of that pays off in search today as well.

Agentic search and off-page SEO

AI agents are likely to turn off-page reputation from a ranking influence into a selection gate. An agent does not hand the user ten results to browse. It reads sources, compares options and comes back with a shortlist or a finished task. What independent sources say about you then decides whether you are in the set at all.

That is a reasoned expectation, not a finding. No reliable public data yet measures how agents choose between brands. This lesson separates what has been announced from what is inferred, and ends with the preparation that makes sense either way.

What agentic search means

In classic search, a person types a query, reads results and decides. In AI search, the system writes an answer from retrieved sources. In agentic search, the system goes further: it plans several steps, runs its own searches, opens pages, compares what it finds and may act, for example by filling a basket or monitoring a topic over time.

The retrieval step underneath is the same one described in how AI search chooses sources. The difference is who does the reading. A human visitor may never see the pages involved.

What has been announced

Development	Status	Confidence
Google “information agents” running around the clock in Search for AI Pro and Ultra subscribers, with generated layouts and persistent dashboards	Announced at Google I/O on 19 May 2026, for summer 2026	High. Google’s own announcement
Agentic browsers: Perplexity Comet, OpenAI Atlas and agentic features in Chrome	Reported as launched in July 2025 and October 2025	Low. Our research saw only search summaries
Commerce protocols from OpenAI and Google that let agents complete purchases	Reported	Low. Same limitation

One figure you may meet is that 45% of shoppers use AI agents for purchase decisions. It appears in vendor blogs without a traceable source. We leave it out, and you should be cautious with any adoption statistic that does not name its survey.

From ranking to shortlisting

This is the central argument, and it is reasoning, not measurement.

A results page has room for ten or more options, and position nine still gets some clicks. An agent asked to “find me three payroll tools for a ten-person company” returns three. To choose them it has to rely on something, and the available evidence is what third parties say: reviews, comparison articles, community threads and press coverage.

Research on current AI answers points the same way. SparkToro and Gumshoe’s January 2026 test found that leading brands appeared in 55% to 77% of AI responses to recommendation prompts, even though the exact list almost never repeated. A consistent set of names exists. If agents inherit that behaviour, being in the set is the whole contest. Co-citation and co-occurrence explains how that set is thought to form.

For off-page work, the shift is from “how many links point at this page” to “what does the independent web agree about this brand”.

Trust a machine can check

An agent acting for a user needs facts it can verify quickly: who the business is, what the product costs, whether reviews are good. Three kinds of information serve that:

Consistent entity data. The same name, description and details on your site, profiles, directories and knowledge sources.
Structured product and pricing data on your own pages.
Review aggregates on the platforms your category uses.

Whether agents weigh these as described is not established. They are low-cost and already useful for search, which is the case for doing them. Off-page E-E-A-T and reputation covers how reputation outside your site is assessed.

Agents read pages differently

Agents fetch pages in real time and extract what they need. Heavy page furniture gets in the way. Peec AI reported in June 2026 that pages with 60 or more links delivered 33% actual content on an agent’s first read, against 78% for lighter pages. It is one vendor’s figure, and we read it through Peec’s statistics summary.

This is an on-page matter, but it touches off-page work in one respect. The third-party pages that mention you are read the same way. A mention in the opening section of a clean article is more likely to be extracted than one buried in a sidebar of a cluttered page. That is an inference from the Peec figure, not a tested result.

Pressure on the publishers you rely on

If agents read pages on behalf of users, fewer humans visit publisher sites. Fewer visits mean less advertising income for the newsrooms and specialist sites that digital PR depends on. Our research notes journalists already report budget and staffing pressure.

The possible consequences run in two directions. Fewer staffed publications would mean fewer places to earn coverage. And publishers short of revenue may sell more placements, which blurs the line between earned and paid mentions. Neither has been measured.

What probably does not change

Agents still find pages through search indexes, and links still help pages rank in those indexes. Our research’s summary of the trend is that the link as a counted unit is declining, while the link as evidence of a real editorial mention is not. An agent that reads an article naming your brand takes in the mention whether or not the link is followed.

Spam rules also carry over. Google’s spam policies already cover attempts to manipulate generative AI responses, and the methods aimed at agents are the ones described in GEO spam and AI manipulation.

How to prepare without betting on predictions

Each step below is worth doing for today’s search. That is the test for anything you do in the name of agents.

Audit your facts. Check that name, description, pricing and contact details match across your site, profiles and listings.
Find your category’s shortlist sources. Identify the review platforms, comparison pages and communities that AI answers cite for your prompts. The trackers in our AI visibility tools comparison show cited URLs.
Earn a place on them. Reviews from real customers, inclusion in independent comparisons and coverage with original data. Link building for AI visibility sets out the method.
Measure mention rate, not position. Run priority prompts many times and track how often you appear.
Keep key pages light. Put the facts an agent needs near the top, in plain text.
Spread your presence. Citation sources rotate quickly, so do not depend on a single platform.

What we do not know

How any agent chooses between comparable brands.
Whether agents weigh reviews, press coverage and community posts differently from current AI answers.
How many people use agents for decisions that involve money.
Whether agent traffic can be attributed reliably.

When studies appear, check who ran them, how many runs they used and whether the vendor sells a product that depends on the result. Our methodology explains how we treat vendor data.

Where to go next

For the evidence on mentions as it stands today, read brand mentions and citations. Terms such as entity and agentic search are defined in the glossary.