Sales Engineering Is the New Marketing: When 94% of Buyers Fact-Check You

Sales engineering used to begin after the demo request. It now begins before it. 94% of technology buyers who use AI to research a purchase go on to fact-check what the AI told them. The fastest way to survive that check is a product a stranger can operate without asking your permission. Messaging still matters. It no longer decides who gets onto the shortlist.

What "sales engineering is the new marketing" actually claims

Sales engineering is becoming a marketing function because the proof it produces now reaches buyers before any seller does. 67% of B2B buyers prefer a rep-free experience. The demo environment, the public benchmark and the reproducible test have become discovery assets rather than closing tools, and they are read by machines as often as by people.

That is the whole argument, and it is narrower than the slogan suggests. Sales engineering is the work of proving that a product does what the brochure says, usually in front of a sceptical technical buyer. Marketing is the work of getting considered in the first place. Those two jobs used to sit at opposite ends of the funnel.

They have collided because the gap between a claim and its verification collapsed. A buyer who reads "sets up in under a day" can now ask an assistant to find evidence, pull your documentation, check your changelog and read three reviews inside a minute. The claim and the audit of the claim arrive together.

What this does not claim is that copywriting stopped working. Positioning still decides which category you get compared inside, and that decision still happens in prose. The change is that prose alone no longer survives contact with the evaluation stage.

The verification layer buyers built in eighteen months

Four independent surveys, run by three organisations with different commercial interests, point the same way. The interesting part is not that buyers use AI. It is what they do immediately afterwards.

Discovery moved to the chatbot

G2's April 2026 research found that 51% of B2B software buyers start research with AI chatbots more than with Google, up from 29% in April 2025. In the same study, 69% said AI guidance led them to a vendor other than the one they had planned on, and 33% bought from a vendor they had not previously heard of.

That second number is the one worth sitting with. A third of purchases went to a company the buyer could not have named at the start. Nothing in a brand campaign produced that outcome, because the buyer never saw the campaign. What produced it was whatever an answer engine could find, read and repeat about the product. The mechanics of that are covered in the piece on what the evidence says about citation in AI answers.

Trust did not move with it

Adoption ran far ahead of confidence. G2 found that 64% of buyers encounter inaccurate AI recommendations often or very often. TrustRadius found that 94% of AI users fact-check the answers at least some of the time. In the same survey, product demos, free trials, prior experience and user reviews remain among the most influential resources for choosing a vendor.

So the buying process now has a research stage that is fast, cheap and machine-mediated, followed by a verification stage that is slow, manual and evidence-hungry. Marketing owns the first. Sales engineering owns the second. The second is where the deal is decided, and G2 puts evaluation at 40% of the buying journey against 36% for research.

All four are self-reported buyer surveys with panels weighted toward North America and toward larger software purchases. None of them measures what buyers did, only what buyers say they did.

Why messaging depreciates faster than proof

A marketing claim has always been a promise about the future. The reader accepted it on credit, because checking it cost more than it was worth. That credit has now been withdrawn, because checking costs almost nothing.

Think about what happens to a claim under those conditions. If it is true and verifiable, the verification step makes it stronger, since the buyer now believes it rather than merely reading it. If it is true but unverifiable, it survives as noise. If it is false, the check finds the gap and you are removed from a shortlist you never knew you were on.

That is an asymmetry, and it has a practical consequence. The expected return on an unverifiable claim has fallen toward zero, while the return on a verifiable one has risen. I would stop spending on the first category entirely and move that budget into the second, which almost always means moving it toward engineering time rather than toward more content.

There is a second-order effect that gets missed. Verifiability is a property of the asset, not of the sentence. You cannot make "fastest in category" checkable by rewording it. You make it checkable by publishing the test. That is why this shift lands on sales engineering rather than on the copy desk, and why case studies with verified outcomes now behave more like evidence than like collateral.

Demo environments are a top-of-funnel asset now

The demo used to be a scheduled event with a person in it. For most software categories it is now a page, and the page is where the evaluation begins rather than ends.

Vendor platform data gives a directional sense of the scale. Navattic, an interactive demo vendor, reports analysing more than 40,000 demos built on its platform in the past year, up 43% on the prior year. Engagement rates in its top cohorts sit between 53% and 56%. Treat those figures as what they are. They are a supplier's data about its own customers, useful for direction and unsuitable as the sole basis for a decision.

The structural point does not depend on that data. If 67% of buyers prefer no rep, and 83% of TrustRadius respondents shortlisted three or fewer products, then the artefact that gets you into a three-name shortlist has to work unattended. A demo behind a form is not an unattended artefact. It is a lead-capture mechanism wearing a demo costume, and the trade-offs of that choice are worked through in the piece on what gating actually costs you.

What a demo has to settle before the first call

The useful test is not whether the demo looks good. It is whether the demo closes a specific question the buyer would otherwise resolve by asking an assistant, which will answer from whatever it can find.

The middle column is the honest part. In four of these five rows the buyer resolves the question without contacting you, and in three of them the default answer is scepticism.

Public benchmarks are the expensive version, and the credible one

A benchmark is the strongest proof asset available to a software company. It is strong precisely because it is costly and risky to publish. Anyone can assert speed. Publishing a test that a competitor can rerun on your numbers is a different kind of statement.

ClickHouse runs the clearest example in enterprise infrastructure. ClickBench publishes results for more than 60 database systems using 43 queries on a 100-million-row dataset, on a stated default machine. The project states that any result can be reproduced in around 20 minutes. The company that maintains it also competes in it.

The same pattern runs at industry scale in machine learning. MLCommons published MLPerf Inference v6.0 on 1 April 2026 with submissions from 24 organisations, including three first-time submitters. Those submissions are marketing. They are also a standing invitation to be measured next to everyone else on the same workload.

What makes a benchmark citable

Four properties, and a benchmark missing any one of them is a chart rather than evidence. The workload has to be public. The hardware and configuration have to be named. The results have to be reproducible by a third party in a stated time. And the maintainer has to publish results where it does not win.

That last property is the one most vendor benchmarks fail, and it is the one an experienced buyer checks first. If every chart in the set ends with your bar on top, the set is not a benchmark. The buyer's assistant will find the competitor rebuttal in the same search that found your page.

The DeWitt clause is why most categories have none

There is a structural reason benchmark publishing is rarer than it should be. Many enterprise software licences carry a DeWitt clause, a term that forbids publishing benchmark results without the vendor's consent. The ClickBench project notes plainly that some vendors do not allow their results to be published for this reason.

If your category is governed by those clauses, a full comparative benchmark may be legally unavailable to you. That does not remove the option, it narrows it. You can still publish a reproducible test of your own system, with the workload, hardware and cost stated, and let a buyer run the comparison in private. Publishing your own numbers under a named configuration is the part nobody can stop you doing, and it sits closer to original research as a content moat than to a comparison page.

The headcount math does not work

If technical proof is now required earlier and more often, the obvious response is to hire more technical sellers. The supply data says that road is closed.

The Bureau of Labor Statistics counted 51,600 sales engineers in the United States in 2025, puts growth at 3% to 2035, and expects about 3,800 openings a year across the whole occupation. That occupational category is narrower than the software industry's use of the title, so read it as a floor rather than a census. Even generously adjusted, it is a small pool growing slowly against an evaluation stage that just expanded to 40% of the buying journey.

Common practice compounds the problem. Typical coverage ratios put one sales engineer against two to four account executives depending on segment, which means technical support is rationed to the deals a manager already believes in. Deals that needed proof in order to become believable never get it.

Turning a sales engineering hour into an asset

The productive framing is not headcount, it is the conversion of hours into artefacts. An hour spent in a live demo serves one account. The same hour spent building a sandbox scenario, a connector document or a reproducible test serves every account that finds it, including the ones that never contact you.

I would put a hard split on this and defend it in the forecast review. At least one day a week of every sales engineer's time goes to assets that outlive the call, protected the way engineering protects on-call time. The objection is always that pipeline coverage suffers this quarter. It does. The alternative is paying for the same hour repeatedly for the rest of the product's life.

Two adjacent decisions follow from the same logic. Evaluation harnesses belong in the same programme, because an accuracy claim is only settled by a test a buyer can rerun, which is the practical case for building your own evaluation suite. And comparison content should be written to be checked rather than to persuade, which is a different brief from the usual one, worked through in the template for B2B comparison pages.

Where this argument is weakest

Three genuine problems, in order of how much they should change your reading.

Buyers still want a human to confirm it

Gartner's May 2026 release found that 69% of B2B buyers turn to a salesperson to validate insights they got from AI. The same release found buyers used an average of seven information sources during a recent purchase. Robert Blaisdell, the Gartner analyst on that work, put it directly. Sales leaders should not read a preference for digital self-service as a signal that sellers matter less.

That materially softens the headline. Proof assets appear to decide who gets considered. People still appear to decide who gets bought. If you cut technical sellers and spent the money on demo software, you would probably lose deals, and I would not run that experiment.

The trust picture is closer than either side of this debate admits. In the same Gartner work, 51% of buyers said they were more likely to meet misleading information from generative AI, against 49% who said the same of a sales rep. Buyers are not switching to AI because they trust it. They are switching because it is faster, and it is barely more suspect than the alternative.

Every figure above is self-reported

All four surveys ask buyers what they did. None observes behaviour. The 94% figure specifically means fact-checking "at least some of the time", which is a much weaker statement than the number looks, and I have seen it quoted without that qualifier repeatedly.

Sample sizes run from about 645 to 1,862, panels skew toward North America and toward larger purchases, and TrustRadius and G2 both sell services to the vendors being written about. That is a Tier 2 evidence base at best. It is consistent across three organisations, which is the strongest thing that can be said for it.

The benchmark argument has its own hole. Public benchmarks get gamed through configuration choice, and a buyer who cannot audit the setup is back to trusting a vendor claim in a more expensive costume. Benchmarks also do not exist in most software categories, and building one for a workflow tool may be impossible rather than merely hard.

Frequently asked questions

Where to start this week

Pick the five claims on your homepage that a technical buyer would most want to check. For each one, write down how a buyer would verify it today without contacting you, and be honest when the answer is that they cannot. That list is your build queue, ordered by how load-bearing the claim is.

Then take the single most-repeated question from your last twenty technical calls and turn the answer into a public artefact this month. A sandbox scenario, a connector document, a small reproducible test. Anything a stranger can operate without permission and an answer engine can read without a login. If nothing on that list can be built in a month, the finding is that your proof is trapped inside people's calendars. The way buyers assemble a shortlist around that gap is set out in the breakdown of the AI-mediated shortlist.

Related reading