AI search is becoming one of several sources a B2B software buyer consults rather than a replacement for the ones they already use, and the near-term shift worth planning for is in measurement rather than in traffic. Buyers are adding AI to an already crowded research process, which means the practical question is not whether they will use it but whether your company survives being summarised by it.
Most writing on this subject is either dismissive or apocalyptic. The measurable position sits somewhere unexciting in between, and that is where a plan has to be built.
What is actually measurable right now
Gartner surveyed 645 B2B buyers between August and September 2025 and found that 45 percent used generative AI during a recent purchase, mainly to gather information about vendors and products. Buyers reported weighing an average of seven different information sources, and 69 percent said they turn to sales representatives to validate what AI told them.
Three things follow from that, and none of them are the headline anybody wants.
AI is additive rather than substitutive. It joined a list of seven sources, it did not shorten the list. Buyers treat its output as a claim to be checked, not as a verdict. And the checking still routes through a human, which means the sales conversation has changed shape rather than disappeared.
The broader adoption picture agrees. The Reuters Institute puts weekly use of AI chatbots for news at 10 percent globally, up from 7 percent a year earlier, with trust at 20 percent against 37 percent for news generally. High use, low trust, rising fast. That combination describes a channel in an awkward adolescence, not a mature one.
The useful framing for a budget conversation. If a buyer weighs seven sources and one of them now summarises the other six, being absent from the summary does not remove you from the process. It removes you from the shortlist that the buyer assembles before the process visibly starts.
Four shifts worth planning for
One. Shortlisting moves earlier and out of sight
The part of the funnel you already could not see is getting larger. A buyer who asks an assistant to compare options in your category is assembling a shortlist without visiting anybody, and the first observable event is a demo request from someone who has already decided who is worth talking to.
The consequence for measurement is severe. Your first-touch data describes the moment somebody surfaced, not the moment they formed a view, and the gap between those two is widening.
Two. Rank stops being the reporting unit
Position tracking survives as a diagnostic and stops working as a report. Answers are composed rather than ranked, they vary between runs, and there is no position to occupy.
What replaces it is duller. A fixed set of buying questions, run repeatedly across engines, counted as a fraction with the denominator stated. Nobody enjoys this metric because it cannot be automated into a pleasing line, and it is the only one that survives contact with how the systems actually behave.
Three. Third-party evidence appreciates
If systems weight independent sources above brand-owned material, and buyers are already checking AI claims against other sources, then the same asset is doing double work. A neutral comparison that ranks you well is now feeding both the summary and the verification step behind it.
The strategic implication is that budget shifts away from publishing volume on your own domain and toward being present, accurately, in the places that get quoted. That is slower, less controllable, and harder to report on, which is why most plans avoid it.
Four. The paid-placement question gets answered
At some point these systems will either carry paid placement or commit publicly to not carrying it, and the answer changes the economics considerably. We do not know which way it goes and neither does anybody selling you a strategy premised on one outcome.
What you can do is avoid building a plan that only works if placement stays unpaid. Earned coverage and a resolved entity keep their value under either outcome. A tactic that depends on the current retrieval behaviour staying exactly as it is does not.
The measurable part of this is available today.
Everything above is directional. What is not directional is whether AI systems can currently reach your site, what they say about you when asked, and which competitors get named instead.
Our free AI visibility check establishes that baseline across four engines and hands back the raw answers. Having it recorded now is what lets you tell a real trend from a bad week later on.
Three things we would not bet on
Forecasts are cheap, so it is worth being explicit about what we think is overstated.
That AI search replaces Google for B2B research. Google still handles the overwhelming majority of search, buyers still use seven sources, and habit is durable. Displacement of the leader is not the near-term story.
That agents will buy software autonomously any time soon. Agent capability is improving quickly, with the 2026 AI Index recording organisational AI adoption at 88 percent in 2025, and procurement of a system that touches contracts and security review is not a task anybody delegates early. Shortlisting, plausibly. Purchasing, not yet.
That any current tactic has a long shelf life. The retrieval behaviour these systems exhibit today is a product decision, not a law. Anything you build that depends on the specific mechanics of 2026 should be treated as rented.
What a 2027 plan should actually contain
Strip out the speculation and a defensible plan is short. Keep your site reachable and renderable by the agents that matter, decided deliberately per agent rather than by a blanket rule. Resolve your entity so that naming you is safe. Establish a measurement baseline now, because you cannot measure a trend backwards. And move a meaningful share of the content budget off your own domain toward the sources that get quoted.
That plan works whether adoption doubles or stalls, which is the only useful test to apply to it.
What changes, by function
Most writing on this topic addresses a generic marketer, which is why it rarely survives contact with a planning meeting. The work lands on different desks depending on which stage is failing, and the split is worth being explicit about before anybody is handed a target.
| Function | What actually changes | What does not |
|---|---|---|
| Demand generation | First-touch attribution loses the AI step entirely, so pipeline appears to arrive from nowhere | The need to be present where buyers form shortlists |
| Content | Value shifts from volume toward passages that answer one question standalone | That useful content still outperforms padded content |
| Product marketing | Category positioning has to be legible to a machine, not just persuasive to a human | The need for a clear answer to which buyer you are for |
| Engineering | Crawler access becomes a marketing-critical setting rather than a footnote | Ordinary technical hygiene, which still covers most of it |
| Sales | Buyers arrive having been told something about you by a third party | That a human closes the confidence gap, per the Gartner finding above |
The uncomfortable one is demand generation. If a buyer reads your name in an answer, does not click, and arrives three weeks later through a branded search, your analytics credits direct traffic and the AI step is invisible. The channel is not underperforming. It is unmeasured, and those look identical on a dashboard.
Leading indicators worth watching
Because the lag between doing this work and seeing anything is long, it helps to track things that move earlier than pipeline does. Four are cheap enough to justify without a business case.
- Branded search volume. The closest free proxy for being mentioned without being clicked. It tends to move before anything downstream does.
- AI crawler hits in your server logs. A direct read on whether the systems are reaching you at all, and the only metric here that reports ground truth rather than inference.
- Citation share against a fixed question set. Tedious, manual, and the only measure that answers the actual question.
- Third-party pages mentioning you. A list, revisited quarterly. It either grows or it does not, and it is the underlying asset behind everything else.
None of those are precise and none require a tool. Together they are enough to distinguish a genuine change from a quiet quarter, which is the only judgement most teams need to make.
The scenario worth stress-testing
Rather than forecasting, it is more useful to ask what would have to be true for your plan to fail badly, and then check whether you would notice.
Suppose adoption keeps climbing, agents get good enough to assemble shortlists reliably, and the systems continue weighting independent sources above brand-owned ones. Under that scenario a company with excellent content and no third-party presence gets quoted as a source and never appears as an option. Traffic looks acceptable. Pipeline quietly degrades. The cause is invisible because nothing broke.
Now check whether your current measurement would catch that. For most companies the honest answer is no, because nothing being tracked distinguishes being cited from being chosen. That gap is worth closing before you need it, and closing it costs an afternoon.
The economics nobody has settled
Underneath the tactical questions sits an unresolved commercial one, and it is worth understanding because it determines whether any of this stays cheap.
Running these systems is expensive. Every answer costs real compute, and unlike a search results page there is no advertising slot attached to it by default. The companies operating them are currently absorbing that cost while they compete for users. That is a phase, not a business model.
Three resolutions are plausible. Subscriptions carry it, which keeps answers unpaid but limits reach to people willing to pay. Advertising arrives, which turns the answer into inventory and changes this discipline into something much closer to media buying. Or the cost falls far enough that it stops mattering, which is the outcome the industry is implicitly betting on.
The reason to care is that two of those three futures make earned coverage more valuable rather than less. If answers carry paid placement, the organic slots get scarcer and the independent sources feeding them get more important. If subscriptions dominate, the audience is smaller and more commercially serious. Only the third outcome, where everything stays roughly as it is, rewards the tactics currently being sold hardest.
Regulation is a real variable, not a footnote
A system that recommends products to buyers is doing something that has attracted regulatory attention in every previous medium it appeared in. Disclosure rules for endorsements exist because a recommendation that looks independent and is not causes measurable harm.
Nobody has yet decided whether an AI recommendation is an editorial output, an advertisement, or something new. If it is eventually treated as an endorsement, disclosure obligations attach to somebody, and the question of who becomes commercially significant for anybody whose visibility depends on it.
We are not predicting an outcome. The planning implication is narrow and worth acting on anyway. Build visibility on things that survive a disclosure regime, meaning accurate information, genuine third-party coverage and a resolved identity. Avoid building on anything that would look bad if the mechanism behind it were published.
What to do in the next ninety days
Forecasting is easy to write and hard to act on, so here is the version that fits in a quarter and does not depend on any of the predictions above being right.
In the first fortnight, settle your crawler policy deliberately across every agent rather than by a single blanket rule, and confirm from your own server logs that the decision is being honoured. This is the only item on the list that can produce a step change, and it is the one most likely to already be broken.
In weeks three and four, write your question set and record a baseline. Ten buying questions, four engines, two runs. It is tedious and it takes one person an afternoon, and without it every claim anybody makes about improvement for the next year is unfalsifiable.
Across the remaining two months, fix the entity if the baseline showed it needed fixing, edit the handful of passages that should be winning and are not, and start one piece of genuinely independent work. That last item is the slow one, so starting it in month three of a quarter rather than month one of next year matters more than how good the first attempt is.
That plan costs almost nothing, produces a measurement you own, and survives every scenario in this article including the ones where the forecasts are wrong.
One thing worth writing down now. Whatever you believe about where this goes, record today what four engines say when asked about your company and your category. In eighteen months that record is the only thing that will tell you whether the change was real or whether the industry talked itself into a trend. Nobody can reconstruct it later, and it costs an afternoon.
Predictions age badly and baselines do not. Of everything in this article, the baseline is the only item that gets more valuable the longer you hold it, which is a strange property for something so dull to have.
Frequently asked questions
Is AI search replacing Google for B2B software research?
No, and the evidence points the other way. Google still handles the large majority of search, and Gartner found B2B buyers weigh an average of seven information sources during a purchase. AI joined that list rather than shortening it.
If buyers verify AI answers with a salesperson, does AI visibility matter?
It matters more, not less. The verification happens after a shortlist exists. Being absent from the summary means you are absent from the list that gets verified, and no amount of sales quality recovers a conversation you were never invited to.
When will AI agents actually buy software without a human?
Not soon for anything touching security review, contracts or procurement. Agents assembling shortlists is already happening. Agents completing purchases in enterprise software is further out than most forecasts suggest.
What should we stop doing?
Reporting AI visibility from a single check, and treating position tracking as the primary measure. Both produce numbers that feel like information and cannot survive being questioned.
How much should a B2B SaaS company spend on this in 2027?
Less than the category is currently pitching, and start with the free diagnostics. Most companies find a technical constraint that costs nothing to fix, and knowing that before signing a programme changes what the programme should contain.
We revise this page as the evidence moves rather than leaving it to age quietly, and we mark what changed. For the underlying mechanics, the complete guide to AEO covers how retrieval and citation actually work.
Shaban runs answer engine optimisation for B2B SaaS companies at Uncited Brands. He works the unglamorous end of the problem, mostly crawler access, retrieval diagnostics and measurement baselines, and publishes the tests that do not go his way alongside the ones that do.