AI Agents as Buyers: Site Readiness for 2027

As of November 2026, more than 50% of internet traffic is no longer human, and daily requests from AI agents have grown 1,700%, according to Search Engine Journal's webinar listing of 11 November 2026. For anyone running ai automation for business, the practical effect is this: part of your buyer research is now done by software that reads your pages, your product data and what other sites say about you, then builds a shortlist before a person ever arrives. The decision for this quarter is not whether to rebuild the site. It is which handful of pages to fix so an agent can read them at all, and web application development time is usually what those fixes cost.

Key takeaways

If half of traffic is already non-human, a page that only renders for browsers is invisible to a growing share of your market.
One enterprise analysis found 96% of a key product page was code AI bots skip, so the agent read a fraction of the file. That is a delivery problem, not a design problem.
Buyers hand over routine research because agents can automate up to 95% of the time such tasks used to take.
Score pages on what an agent can extract, not on how the page looks, and accept that traffic numbers will lag the work.

What changes when an agent does the research

An agent does not browse the way a person does. It fetches a page, strips what it cannot parse, and keeps the text, prices, specifications, availability and any structured markup it finds. Whatever sits in scripts, layout wrappers or image-only blocks disappears before the agent has a chance to compare you with the next supplier on its list.

The volume argument alone should end the debate. If more than 50% of internet traffic is no longer human, and daily requests from AI agents have grown 1,700%, the agent share is not a niche. Search Engine Journal frames the shift plainly: agents compare and shortlist using pages, product data and third-party mentions, and your page design plays no part in that reading.

That reframes how you budget the work. The page has to be readable by a machine before it can be persuasive to the person who reads the agent's summary, and the cheapest reader fix is rarely a redesign.

The page-level readiness scorecard

Advice on this topic tends to stop at "get ready for agents" without saying where to start. Build a scorecard with five checks and score each page zero, one or two. A page that reaches most of the maximum on every check needs no work this quarter.

Check	What you look at	Strong	Weak
Extractable text	View the page with JavaScript disabled, or read its raw HTML	The core offer, price and specifications are in the HTML	Text appears only after client-side rendering
Code-to-content ratio	Compare the size of scripts and markup with the readable copy	Copy dominates the byte count	Almost the whole file is code
Crawler access	Check the robots file and any bot rules	The main crawlers an agent uses are allowed	A blanket rule blocks them
Product data	Look for structured markup on product and service pages	Price, availability and attributes are marked up	Data sits in images or free text only
Third-party description	Search how directories and review sites describe you	Consistent name, category and hours	Contradictory or missing entries

The code-to-content ratio is where the 96% figure lands. In one enterprise analysis, 96% of a key product page was code AI bots skip, so what the agent could use was the remainder. A page like that can rank and still fail at the moment of comparison.

Run the scorecard on the pages that carry the most revenue, not on the whole sitemap. In Singapore, that usually means the product category pages, the pricing page and the contact or enquiry page, where a buyer's agent checks whether a business is reachable and what it costs. The same pass shows where ai automation for business money has been spent on tools that read a page the agent itself cannot read.

Why buyers hand research to agents

The pull side matters as much as the supply side. SynapNews, in its guide for October, reports that for specific operational tasks such as cash reconciliation, menu repricing and initial customer inquiries, AI agents can automate up to 95% of the time previously spent manually. When a buyer has seen that kind of saving inside their own operations, handing a supplier comparison to an agent feels routine rather than risky, and the less the ai automation for business budget is spent on manual reconciliation, the more there is for the site work below.

The same guide records that businesses adopting AI voice agents report a 40-60% reduction in inbound calls for routine inquiries. That number is worth reading twice. It means the first interface a customer meets may be a machine on both ends: their agent asking your agent, or their agent reading the answer off your page.

So the description of your business travels further than your homepage. Yelp's AI assistant accesses 30 days of chat history to keep context, which shows how much weight a memory layer carries in an agent conversation. A brand whose category, service area and price range are stated consistently in every place that memory can draw on is easier to recommend accurately.

How much of the buying decision the agent already carries

Watch what a buyer's agent actually does before you agree to a rebuild. The work it takes over is the comparison step: matching a stated need to a shortlist, checking price and availability, and reading what other sites say. That is why search results about agents keep landing on automation inside the business, while the site side stays thin.

If you build your own agent to answer these enquiries, the same rule applies in reverse. A buyer's agent asking yours for a quote gets nothing usable unless the quote, the service scope and the lead time are written where a machine can read them. Two agents exchanging vague text is a slower version of a person waiting on hold, and getting that handover right is web application development work as much as prompt design.

Which pages need attention first when AI agents do the buying research

Search results on this topic explain what AI agents are and how to deploy them, but they stop short of a sequence for a site owner. The scorecard above is the sequence. Apply it in this order:

Blockers first. Fix any crawler rule that stops agents reaching the site at all. This is an afternoon of work and it gates everything else. A rule written to keep scrapers out often catches the crawlers agents rely on.
High-revenue pages next. Take the revenue pages from the scorecard and fix the ones that fail the extractable-text check.
Product data third. Add structured markup to the pages that lost points on product data, starting with price and availability.
Third-party consistency last, because it needs outreach and takes weeks. Fix your directory entries to match the page.

Treat the order as fixed until the first pass is done. A team that starts by rewriting copy will spend a quarter on words no agent can read.

What you can and cannot measure yet

Attribution is the honest weak point, and the webinar says so when it asks how to measure progress while traffic stays low. Two things are measurable now: the scorecard score per page, tracked monthly, and whether your brand shows up when you put a buyer's real question to an AI engine.

What is not measurable is the revenue an agent influenced before a person clicked. Session-level analytics sees the last visit, not the comparison that happened inside a chat window on someone else's phone. Write that limitation into the report you send upward rather than filling the gap with a number.

Topkee's view: competing for a seat in the answer

The agent-reading-your-pages problem is the one Topkee GEO Service was built around. It monitors whether a brand appears in AI answers to real buyer questions, sampling agreed engines through an API environment and reporting how the brand is described. The method runs six connected workstreams: prompt tracking, query fan-out analysis of the follow-up questions an engine actually returns, sentiment and brand-description accuracy, competitor comparison on the same question set, AI browsing analysis of retrieval sources and cited pages, and knowledge-content writing. That browsing analysis is what tells you which of your pages an engine cites, and it is explicitly not site traffic.

Three monthly tiers carry the work, each capped at a number of concurrently tracked questions and a number of reviewed articles per cycle, so a plan can be sized to how many buyer questions you can answer at once. Three limits belong in any plan you build on it: sampling is not what every consumer sees in a live AI product, AI browsing analysis is not actual traffic, and there is no guarantee of AI recommendation or fixed exposure. Separately, Topkee AI Agent organises professional methods into AI partners that execute tasks and hand back editable files, pages and business objects, which is the build side of the same problem rather than the buyer side.

Recommended next action

Take the scorecard to the pages that carry your revenue and score them this week, then hand the client-side rendering problem to whoever owns the build. If the comparison stage is where you suspect you are losing, look at how non-linear consumer journeys now start inside AI tools before you touch the funnel maths, and keep the scoring honest against AI-assisted campaign measurement rather than against last year's click volumes.

FAQ

Is AI agent traffic really more than half of the internet?

Search Engine Journal's webinar listing, published 11 November 2026, states that more than 50% of internet traffic is no longer human. Treat the figure as a direction of travel from that one source. The non-human share in your own server logs is the number you can act on, and it is usually the first thing to look up before approving a budget.

Which pages should I fix first?

Start with any page where a crawler rule blocks agents, then the pages carrying the most revenue, then product and pricing pages missing structured data, then directory entries that contradict your site. A page that passes the five checks needs no work this quarter.

What does a 96% code-to-content ratio mean in practice?

In one enterprise analysis, 96% of a key product page was code AI bots skip. The agent keeps only the readable remainder, so price, availability and specifications written only in interactive components are lost at the comparison stage even though a browser shows them fine.

Can I tell whether an AI agent recommended us?

Not from web analytics alone. Sampling a set of buyer questions through an AI engine and recording whether the brand appears and how it is described is the workable substitute, as long as you accept that sampling is not what every consumer sees and that no service can guarantee an AI recommendation.

Do we need to rebuild the site?

Usually not. Server-side rendering the revenue pages, adding structured markup and fixing crawler rules covers most of the scorecard. A full rebuild is justified when the platform cannot render text without heavy scripting, and the raw HTML check will show that quickly.

Who should own the fixes?

The build team owns rendering and crawler rules, marketing owns product data and directory entries, and one person should keep the scorecard so the same page is not tested two ways. Escalate the crawler-rule check first because it is the cheapest change with the widest effect.