Half of all business software buyers now start their search by asking an AI model rather than typing keywords into Google, according to G2's own research. A year earlier the figure was 29%.
Most of the money chasing that shift is being spent on the wrong thing. Vendors are rebuilding their own websites to win a pattern match, while the models running a purchase recommendation are doing something different — checking third-party sources to make sure they are not about to cause a purchase the buyer regrets.
"But AI visibility is not the same thing as winning the answer."
Tim Sanders is chief innovation officer at G2, which holds more than three million verified software reviews and is, on its own tracking, the single most-cited source behind AI software recommendations.
The full interview is covered here so you can skip it. 59 minutes of audio, 28 minutes of reading.
Here are the 23 insights that matter.
👤 Guest: Tim Sanders, Chief Innovation Officer at G2, an executive fellow at the Digital Data Design Institute at Harvard Business School and author of "Love Is the Killer App"
🎙️ Host: Craig S. Smith, longtime New York Times correspondent
📰 Published: 14 September 2026 on YouTube (Eye on AI)
🔴 YouTube | 🟣 Apple Podcasts | 🔗 Episode page | ⏱️ 59 min | ✅ Time saved: 31 min
Key Takeaways
Half of B2B software buyers now start in a chatbot rather than a search engine, against 29% a year earlier
Sanders expects roughly two-thirds next year
Being cited by an AI and being recommended by it are different businesses — one makes traffic, the other makes pipeline
On a purchase question the models run a separate validation step, checking third-party authority before they name a vendor
He cites a study putting 41% of the weight on appearance and ranking on authoritative lists
Nothing a vendor writes about itself can win the recommendation
Awards, accreditation and reviews add another 34% of the weight, and none of it can be self-generated
Claude is widely adopted and barely used for software research
He puts ChatGPT and Gemini at 81% of the research use case
Click-through on an ordinary AI citation is a fraction of 1%, but a citation next to a recommendation runs 1–7%
For every dollar spent on answer-engine optimization, companies still spend 60 on search
He predicts third-party agent guardrail services become a $30B industry within five years, at services margins rather than software margins
His closing call is that most successful software companies stop being databases and become harnesses priced on token consumption
1. Knowledge Sharing as Love
Smith opened on the two things in Sanders' background that sit oddly next to a software review company: a fellowship at Harvard and a book about love.
Sanders has spent the last three years as an executive fellow at Harvard's AI institute. "Last three years I've served as an executive fellow at the AI institute at Harvard and the charter of the institute is to democratize artificial intelligence for everyone especially business leaders." It does that through case-level research, publications and executive outreach programs.
The book, "Love Is the Killer App," was his first, written while he was an executive at Yahoo, and its argument was that the more technology there is, the more the human side matters.
"High-tech AI is as high-tech as tech gets, and high-tech leadership and judgment and taste and all those other great human attributes around emotional intelligence have never been more important."
His working definition of love in a business setting is the selfless promotion of the growth of the other — a leader growing a teammate, a company promoting a customer's success.
The mechanism is giving away the three things that grow when shared. "Your knowledge, your network of relationships, and your compassion."
What drew him to the Harvard institute was that its charter is the same move: share knowledge to get businesses off the sidelines on a technology he described as opaque and not yet trusted. He said he now studies trust a great deal.
2. The Trust Layer Mission
Smith put his own framing to him — is G2 the arbiter of software? Sanders answered with the company's name and its founding grievance.
"G2 was founded around the idea of creating the highest form of intelligence for buyers of software." "G2 stands for the highest form of military intelligence." The original company name was G2 Crowd, on the belief that the peers writing reviews are the best source of that intelligence.
The founding grievance was the analyst business. Sanders said the founders were struck by what he called the pay-to-play model of analyst firms like Gartner, and by how hard it was for a software company to establish trust with a buyer.
The reason that matters now is that the models like what G2 has. The frontier labs' agents — he named OpenAI's ChatGPT and Google's Gemini — favor human natural language as expressed in reviews, and G2 has reviews at scale, which he said has made it the dominant player in driving AI search results.
The mission has since been rewritten. "G2's new mission is to become the trust layer for the age of artificial intelligence."
The prize he attached to trust is the size of the gains he says companies are now reporting from agents and autonomous computing. "Not 30%, not 40%, but what we're hearing now, 2x, 5x, 10x in certain situations. Only if they trust it though."
3. Half the Reviews Get Cut
Asked whether the reviews are collected from around the web, Sanders said they are not, and then described how much of what is submitted never appears.
"They're written exclusively on G2's platform." Increasingly they are spoken rather than typed — "We have a voice product now that actually leads to longer reviews."
More than half of what is submitted never reaches the website. The trust and safety process verifies the identity of the reviewer and also that they own the software, which he acknowledged creates a lot of friction.
The friction is the product. He contrasted it with Reddit and other review sites that do not apply that rigor in the interest of collecting the largest number of reviews. "We could have twice as many reviews as we have but we feel like we have enough."
The identity check is also what the models reward — he said the agents love verified identity for establishing trust.
Smith suggested cross-referencing G2's reviews against Reddit's, and Sanders said he has been thinking about it for years. He believes there are thousands, perhaps tens of thousands, of unstructured reviews sitting in communities like Reddit that could be cross-indexed, and that Reddit's upvoting system separates quality from spam even though posters are anonymous.
4. G2, YouTube, Reddit
That led to the ranking Sanders tracks most closely: which sources an AI actually names when it recommends software.
"It's G2, YouTube, Reddit. That's top three right now." G2 is first by a margin; second place alternates between Reddit and YouTube depending on the month.
What is being ranked is citations behind a recommendation, not traffic. His worked example is a commercial intent prompt — a CRM for a medium-sized hospital, month-to-month payments, works well on mobile — where the model returns a short list and names the sources it leaned on.
The scale behind it: Smith put the review count at over three million verified reviews, and Sanders added the rest. "90 million annual visitors and 200 million software buyers globally."
G2 bought three of Gartner's review properties. The acquisition of Gartner Digital Markets — Capterra, Software Advice and GetApp — completed in February 2026. Gartner remains an analyst firm and kept Peer Insights, which its analysts use.
"Capterra's leading software review site, probably top three or top four prior to the acquisition, depending on how you want to count it."
Software Advice is a phone service. "If you're buying software, you can call a 1-800 number and have a conversation with an expert that will help configure for you the exact solution you're looking for."
"And then GetApp really focuses on startups, very small businesses in terms of how they buy software to fuel their companies."
5. What G2 Actually Sells
Smith asked whether the business is subscription or usage-based. Sanders answered both parts.
"Not yet. It's subscription." He added a forecast about the industry generally: "I think the world is going to move to pay for consumption."
The revenue is marketing services, in four pieces. Platform access, bought by claiming your product profiles; reviews as a managed service; buyer intent data, now available through MCP; and advertising.
The intent data is the part he sounded most interested in — not only who is looking at your products, but which of a competitor's customers are looking at them.
His framing of where G2 sits against Google is a stage of the buyer's journey. Google ads are discovery. A review site, an analyst report or a publisher is the evaluation stage, and that is where G2 sells placement.
The pitch to an advertiser is to be added to a short list they did not make, or to disrupt one — including a short list generated by ChatGPT.
6. Half Now Start in Chat
Sanders then gave the behavior number that underwrites the whole business.
"You represent at least half the market according to our latest research." "Like half of all B2B software buyers do it exactly how you do it, Craig. This time last year it was only 29%."
The change is in how the question is asked: natural language to a model rather than keywords to a search engine, and it now comes before anything else in the process.
His forecast for next year: "If I'm looking into the tea leaves next year, probably two-thirds of the people are going to behave like you do, starting in chat, maybe even generating the short list out of chat and then doing the evaluation work."
7. Your Money, Your Life
Smith asked about the industry that has grown up around feeding models positive information. Sanders drew a line between use cases before answering.
Answer engine optimization is a huge industry, he said, but the only use case G2 cares about is the purchase recommendation, because that is what its customers are buying marketing services to win.
"But purchase recommendation represents a unique risk to the user. OpenAI researchers called this your money your life." He put medical advice in the same bucket.
For those prompts the models do not rely on pattern recognition alone — matching the language of the prompt to the language on a website, which he said is exactly what many companies are retooling to do.
What happens instead has a name. Both ChatGPT and Gemini enter what he called validation layer work at inference. "They are validating that the entity that they're about to recommend is not going to lead to a regrettable purchase."
"They look for trust signals from high authority sources, usually with strong underlying data."
The underlying idea is a prediction about your future regret. "I'm trying to predict that if Craig buys this CRM solution, he's going to renew in a year. Because if you buy it and you hate it, that leads to platform health decline." He said the regrettability index dates back to social media and regrettable minutes.
What triggers the extra work is the shape of the prompt. "It's called a commercial intent prompt. It's where the natural language expresses the desire to purchase something."
8. The 41% That Decides
Sanders put a breakdown of the weights on screen, from an outside study.
"One thing I would share is there's a fantastic study out there by a company called First Page Sage." It breaks down how ChatGPT's agent weighs validation at inference time.
The largest single weight is where you appear. He put 41% of the weight on a software company's appearance and, as he put it, its ranking on authoritative lists.
"Another 18% is driven by awards accreditation, 16% by online reviews. All of those are third-party. They can't be generated by a vendor." They must be acquired.
The authoritative lists are of three kinds: review-site lists such as G2's Best Software Awards; lists published by analyst firms like Gartner and Forrester, if those are released for AI crawling; and publisher lists — his example was a technology publication's top ten CRMs for hospitals.
G2 audited where its own citations come from. "Last year, we did an audit for where our citations came from. 60% of our citations came from the best of software awards and all the best of category pages across G2."
He drew a distinction between an authority list and a listicle, saying the models also look at verified identity, sentiment expression and how many people were involved in generating the list.
9. Cited Is Not Recommended
This is the distinction Sanders returned to most often, and the one he says the AEO industry blurs.
"But AI visibility is not the same thing as winning the answer." "Being cited means a link to your content is showing up in an AI return."
The shape of an AI answer is why the two differ. On his CRM example the return is not three names; it is several pages on CRMs in hospitals, mobile endpoints, Android against iOS, pricing structures and month-to-month against flat fee — and only at the end does it make three recommendations.
So content services can get you into the explanation without getting you into the answer. You can generate enough content to show up in the first three pages and pick up a little traffic, he said, and still not win the answer.
"Being cited is how you generate traffic to your website, which hopefully you can convert. Winning the answer generates pipeline, and that's the business we're in at G2."
Asked whether first-party content can do it, his answer was one sentence. "You can't win the answer with vendor claims."
10. Claude Is the Oscars
Smith asked what happens with Claude, which he noted is increasingly used. Sanders separated adoption from use.
Claude is being adopted by a lot of companies, he said, which does not mean their employees research software with it. G2 studies the question every year.
"The answer is Claude is like the Oscars. Everybody talks about the movie. Not many people see the movie."
What it is used for, on his account, is coding, summarizing documents and Slack conversations, weekly automated tasks and — most popular now — building presentation decks. Primary research is not on the list.
He put ChatGPT and Gemini together at 81% of market share for the research use case.
"I would also say Claude has the lowest net promoter score for the research use case of any of the ones I just showed you including Perplexity because it's just not a priority."
His explanation is where a lab spends inference tokens: on what it expects to win. For Claude, he said, that is coding and increasingly design.
One measurable consequence he cited from the answer-engine research firm Profound: on Claude, a research call does live retrieval less than 40% of the time, against 100% on OpenAI. Even with the same algorithm, he said, that changes the results.
A quarter of the buyers are not using corporate tools at all. More than one in four B2B software buyers use their own ChatGPT account on their own phone, because it has their memory: "They don't want to use the corporate one with all the guardrails."
11. Cited Once Means Nothing
Smith asked whether ChatGPT and Claude would give different answers to the same question. Sanders said the bigger problem is that ChatGPT gives different answers to itself.
"Yes. Well, you're likely to get different answers from OpenAI if you ask it today and ask it tomorrow."
The tracking he showed is therefore statistical, not anecdotal. "We're Profound's biggest customer." "They run hundreds of thousands if not millions of synthetic prompts every month to get that statistical body that generates real directionality."
How an AEO engagement works in practice: the firm builds a prompt panel with the client — the questions being tracked — then fans each question out into variants and runs them at volume, averaging the results into a visibility reading.
"On the other hand, if you just went to ChatGPT and put the prompt in to see if your company shows up, it's going to vary every time you do it." Do it enough times, he said, and you reach a statistical average with directionality.
12. The Click-Through Collapse
Asked what the AEO firms are actually doing for clients, Sanders said most of them are chasing traffic rather than pipeline — and gave the numbers that make traffic look thin.
Most companies hire AEO firms to recover the traffic lost to zero-click behavior. G2, by contrast, he said is focused on short lists and shootouts.
The click-through numbers are brutal. Research by PromptWatch put the average click-through rate on an average citation in an AI return at between one-tenth and one-third of one percent.
"Now, if you're cited in relationship to a recommendation, PromptWatch found it's 1 to 7%. It's significantly higher."
What the AEO firms can do, on his account, is make the site easy for machines to consume — the content and the schema structure — so that visibility, mentions and citations go up and some traffic comes back.
The traffic that does arrive is worth more. "Now, what we do know is the traffic you do get from chat bots is higher value traffic because usually by the time they click through to you, they have more conviction." He put it as high as two to three times, and said companies are beginning to separate machine-sourced traffic from human traffic in their reporting.
The direction of travel, though, is down. "But the reality is most companies are realizing organic traffic is going to start getting close to zero as time goes by because we get our answers in the chat and there's no need to poke around."
His explanation for why clicks cluster at the end of a long AI answer is trust. "So now we're at a point where we only click on what is risky to believe on its face, your money, your life." People check the citation next to a recommended purchase and skip the citations in the explanation.
13. Who the AEO Vendors Are
Smith pushed on the mechanics — are these firms spinning up websites across the internet? Sanders said the foundation is your own site, then named vendors by specialty.
"If your website's not optimized to be easy to crawl in training or easy to retrieve at inference, it's kind of game over."
Scrunch, which he said was just acquired by Sitecore, does website remediation — agents that write new code to improve machine readability across a site.
AirOps does content engineering, using agents to find gaps where there is market demand for content a company is not providing. "They call it frontier content if you're first on the scene and you're going to see an outsized number of citations against that content."
Profound runs general-purpose marketing agents — constant reconnaissance, daily reporting on visibility, suggestions or finished content, and schema remediation.
The category itself has changed shape. "I'm seeing a lot of the AEO companies really moving from just dashboards two years ago to always on agents today" to help companies catch up.
Smith's reaction was that he had not seen anything like it for 25 years. "Yogi Berra said it best. It's deja vu all over again."
14. MCP and the Tool Tax
Smith asked about G2's model context protocol capability — how it works and who turns it on.
MCP is a company-wide initiative used internally for G2's own research and automation, and also offered to customers as connectors. Customers are taking up buyer intent data headless, so it lands inside their workflow.
G2 is also building MCP servers with the frontier labs directly, so the agents can consume its verified reviews at inference without friction.
The reason friction matters is economics, not elegance. "And they're not just compute constrained. They are discerning based on how much the user is paying them for inference."
"So I mentioned before if more than one out of four users are using their own ChatGPT at best it's a $20 plan. That's buffet food." Whatever they consume on that plan has to be cheap to consume.
He called MCP the universal USB for connecting data to any system or platform, while keeping an eye on the command line as an alternative.
The trade-off has a name. "So when you use an MCP server and when you call it, it opens up every tool in the shed. Whereas CLI is only going to open up the tool that's needed and called for at the time." He called the overhead the tool tax, and said MCP is still the more general-purpose option in 2026 — and that both beat writing custom APIs every time.
He also corrected a common picture of what a chatbot is doing. "They're not. They're talking to a harness."
"And so the chat when it says retrieving, it's not going on Google and looking something up. It's literally retrieving from a database provided by a vendor who scraped Google" — a step that costs money, and one MCP removes.
From the buyer's side none of this is visible. Asked whether a user has to tell the agent to check G2: "It's seamless. They don't even know it."
15. Agents on a Short Leash
Smith asked how far agents have gone in actually making these decisions. Sanders' answer split sellers from buyers.
"So I think the sellers are using more agents than the buyers."
Buyers keep agents on a short leash: they write a prompt, the agent does the work, they review it, and they evaluate before buying.
The exception is renewals, and it is the number he watches. "However, we're starting to see a trend where companies are telling us not a small number more than 10% in 2026 where agents then are advising if not taking action on routine renewals." That can mean a decision on a contract renewal, an escalation for churn consideration, or an alternative vendor recommendation.
"And I think that's going to be the canary in the coal mine."
His forecast: "I look in the future, Craig, where three years from now, you express business goals and business problems and the agent then will write the prompts, locate the software, and perhaps buy it on your behalf with maybe a couple of checkpoints or guardrails. I see that coming."
"It's not there yet. But for renewals, it's starting to show up."
A worked example of what a seller does with it today: an agent in a co-working tool calls G2 through MCP every week, pulls the data, and triggers downstream actions — a content cadence, or a Slack channel alerting account managers to changes in an account.
16. The Entity Update Lesson
Smith asked how G2 keeps its data authentic at what he called a near-monopoly position. Sanders accepted the premise and then explained why the incentive runs the other way.
"So, yeah, it's true. We probably have 80ish% or more of the market share for reviews."
"Again, we're rejecting more than half of the reviews" to make sure they are not written by bots, generated by vendors through customer conduits, or too overtly recruited.
Over-incentivized outreach gets shut down. Where a vendor runs a program that rewards only good reviews, G2 will unpublish those reviews and may suspend the vendor.
The motive he gave is self-interested. The rigor was originally for the buyer; what the company has learned is that the agent values it, which gives the reviews more weight at inference than reviews from sites without the process.
The cautionary example is what happened to Reddit. "August of last year, 2025, OpenAI institutes the entity update." He described its terms as recommending fewer entities, "We're going to have single ways to talk about companies, and we're going to do validation layer work and really begin to reward real verified human behavior to scale."
"And then we went from like competing neck-and-neck with Reddit for B2B software citations. The game changed overnight."
His read of the standing threat is fake reviews, and he traced it back through Amazon and Google reviews to the birth of review sites. The risk is not the other review sites, he said, but losing the thing that differentiates G2 from analyst firms, publishers and user-generated-content sites.
17. Gatekeeper or Scorekeeper
Smith asked the two uncomfortable questions in a row: does selling marketing services influence the reviews, and will vendors start to see G2 as a gatekeeper?
On influence, his answer was that the service scales volume, not sentiment. "We can help you get more reviews, but you're only going to get like the reviews you deserve." Reviews as a managed service means proactively asking a customer list to write one.
"We're going to publish every review that we capture in a review managed service contract." A one-star review is published.
He said helping a company improve its profile and reputation is not a product G2 sells, and that PR firms and some AEO vendors claim to be in that business. His own view, from 25 years back to AltaVista keyword stuffing: "I think it's a moving target."
On gatekeeping he conceded the point. "anybody that can influence your success is going to be seen as a gatekeeper" — particularly if authority lists carry the biggest weight at test time.
What makes it sharper is who is not in the index. "We know Gartner blocks AI." "A lot of publishers like tier one publishers, AP Bloomberg, they block AI. We don't."
His defense of the line he draws: "That being said, we're not charging you for a review to be put on your website. We're charging you for services that make it easier for you to scale the volume and the recency of your reviews."
Smith called it a fine distinction; Sanders agreed it was a fine distinction, and accepted that a vendor might see today's G2 — a review site with a lot of marketing services — differently from the G2 of a decade ago.
18. A $30B Guardrail Industry
Smith asked which categories in G2's winter 2026 report are moving. Sanders listed the fast ones and then made a prediction about a category that barely exists.
Answer engine optimization is one of the fastest-growing categories, on the amount of disruption in it. Software coding is growing very fast.
Agentic customer service, as distinct from customer service chatbots, is growing fast — his examples were Intercom's Fin, Forethought and Zendesk's agent products.
So is the agent builder category, where companies use platforms such as Agentforce, ServiceNow and UiPath to build large numbers of agents.
AI voice is a breakout category, and orchestration software has exploded — by which he meant full-stack orchestration that brings rules-based AI, robotic process automation and agents into one workflow, not just orchestrating agents.
The prediction is a services business rather than a software one: third-party agent guardrail services, managing guardrails, permission requests, exceptions, rollbacks, remediations and QA for companies whose employees cannot keep up with the risk profile of their own agents.
"They're going to feel more like a services company at 30 to 40%, but that's going to be a $30 billion industry within five years." He contrasted that with a 75% software profit margin.
19. Intent Data at the IP Level
Smith asked about the privacy and ethics of telling vendors which accounts are researching competitors.
The data is signals-based and not addressable. Customers cannot ask G2 to target named users or send messages on their behalf; it is kept at the IP range.
What a vendor sees is company-level. Sanders' example was that people at Nike are researching your product right now, or that a lot of people at Nike are researching your top competitor — but not which department and not which individual, because that is where the privacy covenant with the end user would break.
The distinction he drew against the intent-data incumbents is first-hand against inferred. He described ZoomInfo and 6sense as third-party — inferring behavior by scraping web data — and G2 as second-party, seeing the actual user.
The arms-length treatment is not only a US question, he said; European restrictions set a higher bar again.
He was explicit about what G2 does not build. Asked whether the more targeted product would sell, he said of course people would love to buy it — and then gave a commercial reason for not selling it: run the expected value on the risk and reward and it is not a bankable model for a private company with aspirations to go public. The companies that do offer it, he said, have long been public and are at a scale that can survive the risk.
20. Why Voice Reviews Run Long
Smith asked whether the voice reviews skew young.
Sanders said they do not, twice. What he sees instead is a personality type: "So, voice works the same way, but it seems to cut across demographics." He described it as a productivity profile, people who feel busy, and said a ChatGPT researcher told him the same about ChatGPT voice.
He uses a voice keyboard himself. "I press the function key, I speak, it comes out perfect." He said he is three times faster because he never touches the delete key, the return key or the space bar.
Asked which one, he named it and disclaimed it. "Me personally, this is not a G2 endorsement. I use Willow Voice, but I've used them all."
The business consequence is length. "But what we do know is when they write a voice review, it's longer than if they would have typed it. And it's got more tokens. It's got more content." He said the same is true of prompts entered by voice.
21. Two Rules for Vendors
Asked for the top priorities for a vendor trying to win an AI-first buying journey, Sanders gave two rules and spent most of the time on the first.
"Make yourself easy to work with." Under that heading he gave a checklist:
"So check number one, you need an LLMS.txt file in the root of your website that tells the models what your policies are for crawling and retrieval and that you're open for business."
Block only what information security or critical intellectual property genuinely requires.
"You need to avoid gating. And if you do gate, you need to age gate, meaning gate should disappear on high-value content on the 91st day because most of your leads are generated in the first three months."
"You need to avoid hard to crawl formats like PDF and publish everything you can in HTML or if possible markdown."
"Rule number two is to earn trust signals from dependable lists." Find the high-authority lists and awards that are visible to AI and build relationships with them.
The precondition is the product itself. "You get more reviews because you asked for them. They're good reviews because your product is excellent."
What that changes, in his view, is which growth model is available. "I do think what that does bring up is that we're now in a world where this concept called product-led growth is truly possible that I don't think was possible before AI search." Before AI search, he said, the weight of incumbents' legacy made go-to-market growth the only game in town.
22. $1 on AEO, $60 on Search
Smith asked whether vendors will resent having to pay G2 as well. Sanders reframed the question as a budget problem.
What customers have worked out, he said, is that as more buying starts in AI search they have to invest more in trust signals and depend less on their own site's content — a real paradigm shift from the age of search engine optimization.
The emotional problem underneath it is zero-click behavior, and he was careful about where it started. It was not ChatGPT, he said; it was AI overviews. The resentment he sees is a question about traffic: where is mine and when is it coming back.
"And the answer is it's not. And I think emotionally that's the most difficult thing that companies are having to grapple with."
The spending is still overwhelmingly in the old channel. "For every dollar companies are spending on AEO today, they're spending 60 on SEO tools services and SEM marketing like Google Adwords."
He does not think the reallocation should be dramatic, because the new tools are immature. AEO tools are nascent and it is not clear what works, he said, unlike paid search where a dollar in produces a predictable return — though a smaller one than 25 years ago. "I think it's more like four to 6% this year."
His warning about the old channel was a trivia question about the Yellow Pages, whose best revenue year ever landed at 2007. "So these things go great until they don't."
23. From SaaS to Harness
Smith asked where this ends up — an agentic layer making the decisions, calling pockets of data like G2. Sanders answered with a different prediction and closed the show on it.
He does not expect agents buying on your behalf soon; he called it one of the later use cases. What he does expect is agentic workflows everywhere, because the most successful companies will win by brute force.
The structural change is to the software companies themselves. He quoted Satya Nadella's description of a software-as-a-service company as "a CRUD database with business logic", which he said it has been for 25 or 30 years.
His forecast is that they become harnesses instead, and faster than expected. "What's a harness? Skills, context, governance, connector tools, those four elements. That's a harness."
The business model changes with it. A harness sits on top of language models, frontier or otherwise, and its value-add is a markup on token spend. "They're going to be paid based on consumption. The value of the harness is going to determine their markup."
"I have absolute conviction." His number is that within five years, seven out of ten of the companies successful at that point will have made the switch.
The adoption side moves more slowly. "The way that we as adopters use it will be happening in stages, one use case at a time based on risk profile and trust."
Bonus Insights
Sanders said more than one in four B2B software buyers are good enough at using their personal ChatGPT that they prefer it to the corporate deployment, which is the same behavior that caps what the labs will spend on their inference.
Smith's opening framing of G2 as an impartial arbiter of software went unchallenged; Sanders answered by describing the company's founding idea rather than accepting or rejecting the word.
On whether models are deliberately built this way, Sanders said the validation behavior is fine-tuning applied to particular prompt categories rather than something baked into the base model.
He noted that when an AI is asked to judge a photograph for skin cancer, the same "your money, your life" logic fires — and that the answer in that case is going to be iffy anyway.
On the pay-for-consumption question at the small-business end, he said GetApp and the rest are still subscription, but that he expects consumption pricing to arrive across the industry.
Sanders' bottom line is that the money in AI search has moved from what a company says about itself to what independent sources say about it: the models running a purchase question check third-party authority before they name a name, being cited in the explanation is worth a fraction of a percent of clicks while being named in the recommendation is worth several, and the software companies on the other side of that transaction are about to stop selling seats and start selling harnesses priced on tokens.
Products, Companies & Tools Mentioned
G2 (Sanders' employer: more than three million verified reviews, 90 million annual visitors, and by its own tracking the most-cited source behind AI software recommendations)
Gartner (The analyst firm whose pay-to-play model G2 was founded against, which sold G2 three of its review properties and which Sanders says blocks AI crawlers)
Capterra, Software Advice and GetApp (The three Gartner Digital Markets properties G2 acquired in February 2026 — a top-four review site, a phone service that configures a solution for you, and a startup and small-business site)
Reddit (Top-three cited source for B2B software, and the cautionary tale: its citation share fell away after OpenAI's entity update)
ChatGPT and Gemini (The two models Sanders says run validation-layer work on purchase questions, and together 81% of the software research use case)
Claude (Widely adopted for coding, summarizing and presentation decks, but with the lowest net promoter score for research and live retrieval on under 40% of research calls)
Profound (The answer-engine optimization firm whose synthetic prompts G2 buys at volume; G2 is its biggest customer)
First Page Sage (Published the breakdown Sanders showed, putting 41% of recommendation weight on appearance and ranking on authoritative lists)
Scrunch and Sitecore (Scrunch's agents rewrite a site's code to improve machine readability; Sanders said Sitecore had just acquired it)
AirOps (Content engineering — agents that find content gaps and produce what he calls frontier content)
Willow Voice and Whisper Flow (Voice keyboards; Sanders uses Willow Voice personally and says voice input produces longer reviews and longer prompts)
Intercom's Fin, Forethought and Zendesk (His examples of agentic customer service, one of the fastest-growing categories in G2's winter 2026 report)
Agentforce, ServiceNow and UiPath (The agent-builder platforms companies are using to build agents across their organizations)
ZoomInfo and 6sense (Third-party intent-data firms that infer behavior by scraping; Sanders positions G2's data as second-party)
HubSpot (Has direct connectors into G2's intent data for marketing outreach)
PromptWatch (Source of the click-through figures: a fraction of 1% on an average citation, 1–7% next to a recommendation)
Books & Resources Mentioned
Love Is the Killer App – Tim Sanders (His first book, written while he was a Yahoo executive, arguing that the higher the technology the more the human attributes matter)
The Digital Data Design Institute at Harvard (The Harvard Business School institute where he has spent three years as an executive fellow; he describes its charter as democratizing AI for business leaders)
G2 Best Software Awards (The awards and best-of-category pages that produced 60% of G2's own AI citations last year)
Model Context Protocol (What G2 uses to put its verified reviews and buyer intent data in front of agents; Sanders calls it the universal USB and weighs it against the command line)
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: