TL;DR: Agents browsing on a buyer's behalf is real and happening now. Most of what is written about it is forecast rather than observation, and the forecasts are large. The useful news is that preparing for agents requires almost nothing new: unambiguous pricing, complete specifications, ungated documentation and clear capability statements are the same work that improves ordinary AI visibility.
Key Takeaways
Agents do not start by browsing. Discovery still runs through the same citation sources as any AI answer, so a brand absent there is never a candidate for the browsing stage.
The forecasts are not evidence. Survey findings about intent to deploy are real; predictions about what share of purchasing will be agent-led are projections, and should be cited as such or not at all.
Ambiguity is the failure mode. A human reads around a vague pricing page. An agent comparing specifications treats a missing value as absent.
Nothing on the preparation list is agent-specific. Published pricing, complete specs, open documentation and plain capability statements help human buyers and ordinary AI answers equally.
Blocking is a decision with a cost. Blocking agent traffic blocks the research half of buyers delegate, and the training and retrieval distinction still applies.
The measurement gap widens. An agent reading on a buyer's behalf produces no clicks and no form fills, so self-reported attribution matters more than it already did.
Anchor targets assume your CMS generates slug IDs from heading text. Confirm on staging and adjust the hrefs if your platform slugifies differently.
What is actually happening
This topic attracts large claims. Separating what is observed from what is projected is the first useful thing to do with it.
Observable now
Fact | Note |
|---|---|
Assistants can browse multiple sites and compile comparisons on request | Available to users of the major assistants today |
Enterprises are running agents against routine vendor evaluation | Documented in consultancy research, though scale is unclear |
Commerce protocols for agent-initiated purchase exist | OpenAI's Agentic Commerce Protocol, currently consumer retail |
Browser tooling has begun measuring agent-readability | Google's Lighthouse added an agentic browsing category |
Survey findings, which are evidence of intent rather than behaviour
Deloitte's 2026 enterprise AI survey of several thousand leaders across two dozen countries reported that around three quarters of companies plan to deploy agentic AI within two years.
That is a real finding about stated plans. It is not a measurement of how many purchases agents currently influence, and the two are frequently presented as if they were the same thing.
Forecasts, which are not evidence
Predictions that a given share of B2B selling will involve agent-led negotiation this year, or that most purchases will be agent-mediated within a few years, come from analyst projections. They may prove right. They are not observations and should never be cited as findings.
If you use one, attribute it as a prediction with the firm and year attached.
Why the distinction matters commercially
A programme justified on a forecast is vulnerable when the forecast slips. A programme justified on the observation that agents browse now, plus the fact that preparing for them is the same work that improves ordinary AI visibility, is robust either way.
That second framing is the honest one and it is also the more persuasive one internally.
The practical position
Do the preparation because it pays off today regardless of how fast agents arrive. Nothing in this article asks you to spend on something that only pays off if a prediction comes true.
Agents do not start by browsing
The most important thing to understand, and the one that connects this topic to everything else in the cluster.
The sequence
An agent asked to evaluate vendors does not open a browser and begin crawling your category at random. It first has to decide which vendors to look at, and that decision comes from the same place any AI answer comes from: retrieved sources naming candidates.
Only then does it visit sites, read pricing pages and compare specifications.
What that implies
Stage | What decides it | Your lever |
|---|---|---|
Candidate selection | Retrieved sources naming vendors | Corroboration and citations |
Site visit and extraction | Whether your pages can be read and parsed | Access, structure, specificity |
Comparison | Whether your facts are complete and unambiguous | Pricing, specs, capabilities |
Recommendation | All of the above | All of the above |
A brand absent from the first stage never reaches the second. No amount of agent-readable pricing helps if nothing cited you as a candidate.
The consequence for prioritisation
Agentic readiness is not a separate programme that leapfrogs your visibility work. It sits on top of it, and the work underneath is unchanged: be reachable, be quotable, be corroborated.
Teams treating agent preparation as a new initiative while their citation rate is near zero are optimising a stage they never reach.
What is genuinely new at the browsing stage
Once an agent is on your site, it behaves differently from a human. It does not scroll past a vague pricing section and form an impression; it extracts fields and finds one missing. That difference is real and it is what the rest of this page covers.
The connection worth stating internally
If you are already doing the citation and corroboration work, you are already doing most of agentic preparation. The remainder is a legibility exercise on a handful of pages, which is a much smaller ask than the topic's framing usually suggests.
What breaks for an agent
A human and an agent fail differently on the same page, and the agent's failures are harsher.
Your page | A human does | An agent does |
|---|---|---|
Pricing says contact sales | Forms a rough impression, maybe enquires | Records no price, excludes you from any budget-constrained comparison |
A capability is implied in marketing copy | Infers you probably do it | Finds no explicit statement, treats it as unconfirmed |
Specifications are in a downloadable PDF | Downloads it if motivated | Frequently does not, and the data is absent |
Integrations shown as a logo grid | Recognises the logo and assumes it works | Cannot confirm what the integration does |
Key content requires a login | Signs up if interested enough | Stops |
Content renders client-side only | Sees the page normally | May see nothing |
A comparison table is an image | Reads it | Cannot extract it |
The pattern
Humans fill gaps with inference and benefit of the doubt. Agents comparing on specific criteria treat an absent value as absent, and an absent value on a required criterion is an exclusion.
That is the whole difference, and it explains why the preparation list is mostly about removing ambiguity rather than adding persuasion.
The harshest case
A required capability that you have, stated nowhere explicitly. The agent excludes you, the buyer never learns you were a candidate, and you never learn the evaluation happened.
That failure is silent and it is the same failure as a false negative in an ordinary AI answer, which is covered elsewhere in this cluster as the most urgent thing to fix.
What does not break
Brand, design, narrative and positioning still matter, because a human approves what the agent recommends and makes the final decision. Agents narrow; people choose.
So this is not an argument for stripping persuasion out of your site. It is an argument for making sure the facts are also there, extractably, alongside it.
The quick test
Take your pricing page, your main product page and your integrations page. For each, list the facts a buyer with a specific requirement would need. Then check whether each one is stated in text, on the page, without a download or a login.
The four phases
Agent-mediated evaluation runs in stages, and each needs something different from you.
Phase | What happens | What it needs from you |
|---|---|---|
Discovery | Candidates assembled from retrieved sources | Citations and corroboration. Nothing agent-specific |
Qualification | Candidates checked against hard requirements | Explicit capability and compatibility statements |
Evaluation | Shortlisted options compared on specifics | Complete specifications, unambiguous pricing |
Approval | A human decides from the agent's shortlist | Everything persuasion has always done |
Where most companies fail
Qualification. It is the cheapest stage to fix and the one nobody has prepared for, because the facts exist internally and are nowhere stated plainly.
Does it support single sign-on. Does it integrate with a particular system. Is it available in a given region. Does it hold a given certification. Each has a yes or no answer you know, and most sites imply rather than state them.
The approval phase is why brand still matters
A human approves what the agent recommends, partly on credibility built through content the agent itself does not evaluate. Case studies, reputation and positioning do work at this stage.
So the correct reading is not that marketing becomes irrelevant. It is that marketing now has to clear a factual filter before it gets the chance to persuade.
The stage-by-stage audit
For each of your top requirements, ask where the answer lives.
Requirement type | Where it should be |
|---|---|
Capability | A sentence on the product page, in plain text |
Compatibility | An integration page per major integration |
Compliance | A security or trust page, with certifications named and dated |
Commercial | A pricing page with numbers or an explicit structure |
Regional | Stated availability by region |
What this adds up to
Mostly pages the B2B SaaS guide already recommends for human buyers. The agent case sharpens the argument rather than changing the work, which is the honest summary of this entire topic.
Blocking agents
A real decision with real trade-offs, and it is being made by default at most companies rather than deliberately.
The three categories, again
Category | Blocking it costs you | Blocking it protects |
|---|---|---|
Training crawlers | Nothing in AI search visibility | Your content from training use |
Retrieval crawlers | Removal from AI answers | Nothing much |
User-triggered agent fetches | The research your buyers delegated | Server load, scraping |
The third is new and is what agentic browsing introduces. It is a person's agent fetching your page because that person asked it to, which makes blocking it closer to blocking a visitor than blocking a scraper.
The honest tension
Agent traffic has real costs: load, no advertising value, no session to measure, and genuine abuse potential. Security teams have legitimate reasons to challenge automated traffic.
Against that, blocking agent fetches blocks part of a buying process that increasingly happens without a human opening your site at all.
How to decide
Separately per category, with the right people in the room. Treat training as a licensing question for legal and communications, retrieval as a marketing visibility question, and user-triggered fetches as a commercial question about whether you want to be evaluable.
What does not work is one blanket bot rule, which is what most organisations have.
Verify rather than assume
As always, server logs rather than robots.txt. Filter by the agent user agents and look at status codes. A permissive robots.txt with a bot manager returning challenges produces exactly the symptom nobody can explain.
The thing to revisit
Many organisations blocked broadly during the opt-out wave a few years ago and have never revisited it. Those rules were often written before the distinction between training, retrieval and user-triggered fetching was clear.
A deliberate review now, with the three categories separated, is worth an hour and frequently finds a block nobody intended.
Pricing and specifications
The two places where ambiguity costs most, because they are the two things comparisons are built on.
Pricing
An agent filtering on budget needs a number. Contact sales yields nothing to filter on, so you are either excluded or carried forward with an unknown that a later stage resolves against you.
What to publish | Why it works |
|---|---|
A starting price | A floor to filter against |
Ranges by tier | Enough to answer is this in budget |
What drives the price | Seats, usage, modules. Explains variability honestly |
A worked example | For a team of this size, expect roughly this |
An explicit enterprise note | Quoted individually, with the factors listed |
The last row matters. An explicit statement that enterprise pricing is bespoke, with the factors named, is information. Silence is not, and silence is what most pages offer.
Specifications
Complete, in text, on the page. Not in a datasheet PDF, not in an image, not implied.
Include | Because |
|---|---|
Hard limits and capacities | Common filter criteria |
Supported standards and protocols | Yes or no questions with real answers |
Deployment options | Cloud, on-premise, hybrid, regions |
Authentication and access control | Frequently a hard requirement |
Integration detail, per integration | What syncs, which way, how often, what does not |
What you do not support | The rarest and most useful statement |
The negative statement
Stating what you do not do seems commercially counterintuitive and is unusually valuable. It is specific, it is quotable, and it prevents bad-fit evaluations that waste everyone's time.
It also makes everything else on the page more credible, for an agent assembling a comparison and for a human reading it.
Keeping it current
Specifications go stale with every release. Whatever process updates the product also needs to update these pages, or you end up precisely wrong rather than vaguely right, which is worse.
Date them visibly. It costs nothing and it gives a retrieval system a reason to prefer your current page over an older third-party description.
Programmatic access
The one genuinely new item, and the one to be most careful about recommending.
The argument
If an agent can query your product's data directly rather than scraping a page, it gets accurate current information and can act on it. Public API documentation, machine-readable product data and, in commerce, participation in an agent checkout protocol all sit here.
There is a reasonable case that this will matter more over time, and some evidence it already functions as a signal in itself.
What exists today
Thing | State |
|---|---|
Public API documentation | Standard practice, and useful to agents as to humans |
Structured product data on your pages | Standard schema, worth doing |
Commerce protocols for agent purchase | Real, consumer retail, by application |
Agent-readable capability declarations as a standard | Not settled. Proposals exist; nothing widely adopted |
That last row is where caution belongs. Several proposed formats circulate and none has the adoption that would make implementing it obviously worthwhile.
What I would do and not do
Do | Do not |
|---|---|
Keep API documentation public and current | Build against an unadopted standard because an article recommended it |
Use standard structured data properly | Add speculative metadata files |
Publish machine-readable pricing where you can | Rearchitect for agents on a forecast |
Watch the commerce protocols if you sell online | Assume B2B equivalents exist yet |
The deprecation risk
This area will produce several formats, most of which will not last. Work done against a format that does not get adopted is wasted, and the cost of waiting is low because the underlying preparation, clear facts in readable pages, serves agents adequately in the meantime.
The exception
If you sell consumer products online, the commerce protocols are live and real and the ecommerce guide covers them. That is a current decision rather than a forward-looking one.
For B2B and services, the equivalent infrastructure is not settled, and clear pages are the right investment.
The measurement gap
Agentic evaluation widens a problem the cluster already treats as structural.
What disappears
Signal | What happens |
|---|---|
Session behaviour | An agent's fetch is not a browsing session |
Time on page | Meaningless |
Form fills during research | The agent does not fill forms on the buyer's behalf, usually |
Click-through from a recommendation | The human may never click at all |
Any trace of an evaluation you lost | None whatsoever |
That last row is the serious one. A buyer's agent can evaluate you, exclude you on a missing specification, and produce no record anywhere that it happened.
What still works
Signal | Note |
|---|---|
Prompt-set measurement | Still the primary instrument. Ask the questions yourself |
Server logs | Agent fetches appear here, by user agent |
Self-reported attribution | More important than ever, because it is the only route to the invisible cases |
Branded search | Still a directional proxy for influence without a click |
The log signal worth watching
If user-triggered agent fetches start appearing in your logs for product and pricing pages, that is evidence of evaluation happening. It is crude and it is a genuine signal that this is live in your category.
Worth adding to whatever log analysis you already do for crawler access, since it is the same filter with different user agents.
The intake question to add
Ask new customers whether an AI assistant or agent was involved in their research, and if so, what it did. That question is already recommended in the attribution article; agentic evaluation raises its value because it covers cases with no other trace.
The honest reporting line
We cannot see evaluations we lost, and agentic research increases the share of evaluations we cannot see. Our measured figures were always a floor, and this widens the gap between the floor and reality.
That is uncomfortable and it is accurate, and stating it is better than discovering later that a confident number omitted an entire category of buying behaviour.
What to do now
The test for anything on this page: would it be worth doing if agentic buying stalled for three years? Everything below passes.
Worth doing now
Action | Pays off regardless because |
|---|---|
State capabilities explicitly in plain text | Humans and ordinary AI answers need this too |
Publish a pricing position | Already the single most common gap for human buyers |
Complete specifications in text on the page | Helps every reader, machine or not |
An integration page per major integration | Answers a question buyers ask constantly |
Ungate documentation | Your most quotable content, currently invisible |
Review bot rules across the three categories | You may be blocking retrieval already |
Add the agent question to intake | Costs one form field |
Worth watching, not acting on
Thing | Why wait |
|---|---|
Agent-readable capability formats | No settled standard. Early work may be wasted |
B2B commerce protocols | Not established as the consumer ones are |
Agentic readiness scores from vendors | No agreed definition. Ask what is counted |
Restructuring your site for agents | Premature, and the facts work is the real content of it |
The sequencing argument
If your citation rate is low, that is the constraint, because an agent never reaches your site if nothing named you as a candidate. Fix discovery first.
If your citation rate is healthy and your pricing page says contact sales, the legibility work is the constraint, and it is a fortnight of effort.
How to frame it internally
Not as preparing for agents, which invites a debate about forecasts you cannot settle. As making our facts legible, which is defensible on today's evidence and happens to also prepare you.
That framing gets approved. The forecast framing gets argued about.
The one thing to do this week
Take your three most important product pages and list the facts a buyer with hard requirements would need. Check each is stated, in text, without a download or a login.
The gaps you find are the work, and most teams find more than they expect.
Six mistakes
1. Citing forecasts as findings. Survey data about intent to deploy is real. Predictions about what share of buying will be agent-led are projections. Presenting the second as the first is the most common error in this topic and the easiest to be caught on.
2. Treating agentic readiness as a separate programme. Agents do not start by browsing. If nothing cites you as a candidate, your agent-readable pricing page is never read.
3. Building against unsettled standards. Several agent-readable formats are proposed and none is widely adopted. Work done against one that does not survive is wasted, and clear pages serve adequately meanwhile.
4. Blanket bot blocking. Training, retrieval and user-triggered agent fetches are three different decisions with three different costs. Most organisations have one rule covering all three, written before the distinction was clear.
5. Leaving capabilities implied. A human infers from marketing copy; an agent comparing on criteria records an absent value. A capability you have but never state plainly is a silent exclusion.
6. Assuming brand stops mattering. A human approves what the agent shortlists, partly on credibility the agent never evaluated. Agents narrow and people choose, so persuasion still has a job, after the factual filter.
The pattern
Two are about over-reacting to the topic, two about under-reacting, and two about getting the mechanism wrong. That spread is typical of a subject where the commentary runs well ahead of the evidence.
The one that costs most
Number three. It is the only item here that can waste real engineering time on something that may never matter, and the pressure to do it comes from articles with a commercial interest in urgency.
The facts work cannot be wasted. That asymmetry is the whole argument for doing the unglamorous version first.
FAQs
What is agentic search?
An AI agent researching, comparing and shortlisting on a buyer's behalf rather than the buyer browsing themselves. The agent reads the pages, extracts the facts and returns a shortlist, which a human then approves.
Is this actually happening or is it hype?
Both. Assistants can browse and compile comparisons today, and enterprises are running agents against routine vendor evaluation. The larger claims about what share of purchasing will be agent-led are analyst forecasts rather than measurements.
Do agents browse our site directly?
Only after deciding you are a candidate, and that decision comes from retrieved sources naming vendors. A brand absent from that stage is never visited, which is why ordinary AI visibility work comes first.
What breaks when an agent reads our site?
Ambiguity. A human infers from marketing copy and reads around a vague pricing page. An agent comparing against specific criteria records a missing value as absent, and an absent value on a required criterion is an exclusion.
Should we block AI agents?
Decide separately for training crawlers, retrieval crawlers and user-triggered agent fetches. The third is a person's agent acting on their request, so blocking it is closer to blocking a visitor than blocking a scraper.
Do we need special agent-readable files?
Not yet. Several formats are proposed and none is widely adopted, so work against one may be wasted. Clear pricing, complete specifications in text, and ungated documentation serve agents adequately and help human buyers today.
Does brand still matter if agents do the shortlisting?
Yes, at the approval stage. A human chooses from the agent's shortlist, partly on credibility built through content the agent did not evaluate. Agents narrow the field; people decide. The factual filter comes first.
How do we measure agentic evaluation?
Mostly you cannot. An evaluation you lost leaves no trace. Watch for agent user agents in server logs, keep running your own prompt set, and ask customers at intake whether an agent was involved in their research.
Getting ready without betting on it
Take your three most important product pages and list the facts a buyer with hard requirements would need: capabilities, compatibility, limits, compliance, price. Check each is stated in text, on the page, with no download and no login.
The gaps are the work. None of it depends on any forecast being right, which is the point.
The order
Discovery first. If nothing cites you as a candidate, agents never reach your site.
State capabilities explicitly. The cheapest fix and the most common gap.
Publish a pricing position. A number, a range, or an honest explanation of the model.
Complete the specifications, in text, dated.
Review bot rules across training, retrieval and user-triggered fetches separately.
Add the agent question to your intake form.
And watch, without building: agent-readable formats, B2B commerce protocols, and whatever Lighthouse-style tooling emerges for measuring this.
If the production is the constraint
Most of the list above is pages: capability statements, integration pages, specification pages, a real pricing page. Producing those consistently is where teams stall, and Blazly SEO covers research, planning, drafting and direct publishing from about $39 a month.
The factual accuracy stays with you, and on specification and compliance pages it matters more than usual.
Summary
Agents browsing on a buyer's behalf is real now; most of the claims about its scale are forecasts and should be cited as such. Agents do not start by browsing, so discovery still runs through the same citations as any AI answer and a brand absent there is never visited. What changes at the browsing stage is that ambiguity becomes exclusion: a human infers from marketing copy, while an agent comparing on criteria records a missing value as absent. The preparation is therefore unglamorous and familiar: explicit capability statements, published pricing, complete specifications in text, ungated documentation. None of it is agent-specific and all of it helps human buyers today, which is why it is worth doing whether or not the forecasts land. Hold off on building against unsettled agent-readable standards, decide your bot rules deliberately across the three crawler categories, and accept that evaluations you lose to agents leave no trace at all.