The Trust Layer Behind AI Software Recommendations: Tim Sanders on G2, AEO, MCP, and Agentic Buying

1. Guest Background

This episode of Eye on AI, hosted by Craig S. Smith, features Tim Sanders of G2 and takes its title from the central question of the interview: what hidden algorithm decides which software AI will recommend? At nearly an hour, the conversation is less a product announcement than an extended analysis of B2B software discovery, review platforms, validation layers, answer-engine optimization, MCP, and the slow movement toward agentic purchasing.

Sanders first situates himself through two roles. He says he joined G2 in the fall of 2024 and now serves as chief innovation officer. Before that, he spent three years as an executive fellow at Harvard's AI Institute. As he describes it, the institute’s charter is to democratize AI, especially for business leaders, through case-level research, publications, and executive outreach programs. That background matters because Sanders frames AI adoption less as a pure capability problem than as a trust problem: leaders may understand that AI is powerful and still hesitate if they cannot judge when to rely on it.

Smith also brings up Sanders’ book Love Is the Killer App, which might seem far removed from software recommendations until Sanders connects it to trust. In the interview, he defines love in a business context as the selfless promotion of another person’s growth, and he names knowledge, networks of relationships, and compassion as assets that grow when shared. That idea becomes a quiet through-line for the episode: in software buying, the most valuable information is not a vendor’s self-description but evidence that helps another buyer avoid a bad decision.

G2 is presented as a software review and buyer-intelligence platform that began as G2 Crowd. Sanders says the founding idea was that the crowd, reviews, community, and peers could generate a higher form of software-buying intelligence than the pay-to-play analyst model. In the AI era, he gives G2 a more ambitious mission: to become the trust layer for the age of artificial intelligence. His argument is that agents and autonomous computing may produce 2x, 5x, or 10x business outcomes in some settings, but only if buyers and users trust the systems enough to act on them.

2. What the Episode Covers

The main arc of the episode is the migration of B2B software discovery from search-engine results pages into model conversations. Sanders says G2 research finds that roughly half of B2B software buyers now ask an AI model first instead of beginning with Google keywords; he says that figure was 29% a year earlier and predicts it may reach roughly two-thirds next year. Those numbers are Sanders-cited claims from the episode, not independently established market facts, but they explain why G2 now cares about how models consume review evidence, not only how humans read review pages.

G2 is not tracking every AI conversation. Sanders focuses on commercial-intent prompts, where the user’s language expresses a desire to buy, compare, shortlist, or evaluate software. His example is a buyer asking for CRM software for a mid-sized hospital, on month-to-month payments, with strong mobile support. When the model returns a shortlist, it cites sources behind the recommendation. Sanders says G2 tracks that influence and is currently number one by a margin, while Reddit and YouTube alternate for second place depending on the month.

The episode also covers G2’s scale and business model. Sanders says G2 has more than 3 million reviews, 90 million annual visitors, and 200 million global software buyers. He also says G2 completed the acquisition of Gartner Digital Markets properties Captera, Software Advice, and GetApp in February 2026. G2’s revenue model, as he describes it, is subscription-based marketing services: vendors claim product profiles, use managed review services, access buyer intent data, make data intelligence available through MCP, and advertise in the evaluation stage of the buyer journey.

All of this leads to one of the episode’s most important distinctions: AI visibility is not the same as winning the answer. Sanders says being cited in an AI response can generate traffic, but appearing in the final recommendation shortlist is more likely to generate pipeline. He cites PromptWatch research saying ordinary citation click-through rates average around 0.1% to 0.33%, while recommendation-related citations can reach about 1% to 7%. He also says G2 received 1 million human visitors from OpenAI userbot last year. These numbers should be read as claims relayed in the interview rather than universal benchmarks.

The second half of the episode broadens into AEO, MCP, review governance, and automated buying. In Sanders’ framing, AEO is not simply keyword stuffing for chatbots. It includes making content easier to train on, retrieve, cite, and validate. He discusses prompt panels, synthetic prompts, Scrunch, AirOps, and Profound as examples of tooling around visibility measurement, machine-readability remediation, content engineering, reporting, artifacts, and schema remediation. MCP is G2’s route into agent workflows: internally for research and automation, for customers through headless buyer-intent workflows, and for Frontier Labs so agents can consume G2 verified reviews during inference.

3. Core Views: Reasoning, Examples, and Limits

Sanders’ first major view is that software purchase recommendations are different from ordinary information retrieval. A generic lookup may depend heavily on pattern recognition across web text, but purchase recommendations carry real cost, renewal risk, and regret risk. He says ChatGPT and Gemini therefore perform validation-layer work at inference for commercial-intent prompts, looking for high-authority trust signals with strong underlying data. That reasoning explains why vendor website content, even when crawlable, may help a company be mentioned without making the model confident enough to recommend it.

He uses First Page Sage’s analysis as a concrete example. According to Sanders’ citation of that work, ChatGPT software recommendation validation weights include 41% for appearance and ranking on authoritative lists, 18% for awards or accreditation, and 16% for online reviews. He also says G2’s own audit found that 60% of its AI citations came from Best of Software Awards and Best of Category pages. The limitation is important: these figures are cited in the episode by Sanders and should not be generalized into a permanent law of AI recommendation. They are strongest as evidence for how G2 understands its own strategic position.

His second view is that G2’s moat is not merely the volume of reviews, but whether models treat those reviews as verified human behavior. Sanders repeatedly says G2 reviews are written on G2, that the company verifies identity and software ownership, and that more than half of submitted reviews do not publish. Later in the interview, he adds that G2 fights bot reviews, vendor-generated reviews through customer conduits, and programs that over-incentivize positive reviews; violations can lead to unpublished reviews or vendor suspension. Because models may assign more weight to verified reviews, review integrity becomes not just a platform-quality issue but a distribution issue inside AI recommendations.

That same logic creates a platform-power problem. Sanders acknowledges that vendors may see G2 as a gatekeeper, especially if high-authority lists matter at test time and if some publishers or analyst firms block AI crawling while G2 does not. He also says G2 may have around 80% or more of the software review market. Sanders’ defense is that G2 does not charge to place reviews on vendor websites or to make reviews positive; it charges for services that increase review volume and recency. Still, he understands that vendors may see the modern G2 as a review platform with more marketing services and greater influence than before.

His third view is that data-source usability now matters almost as much as data-source authority. Sanders argues that when users think they are talking directly to a model, they are often interacting with a harness; when the interface says it is retrieving, it may be pulling from a vendor-provided scraped database rather than visiting Google itself. MCP matters because it can lower the cost and friction of real-time retrieval and make a trusted data source easier for agents to consume. Yet he also names a tradeoff: MCP has a tool tax because calling a server can expose all tools, while a CLI may call only the tool needed. MCP is therefore not magic; it is a connector strategy with tradeoffs among generality, cost, tool exposure, and workflow convenience.

His fourth view is that agentic procurement is coming, but not all at once. Sanders says sellers appear to be using agents more actively than buyers today, and buyers often keep agents on a short leash: prompt, review the work, then decide. The early signal he highlights is routine renewals. In G2 research, he says more than 10% of enterprises in 2026 report agents advising or taking action on routine renewals, including renewal decisions, churn-escalation considerations, or alternative vendor recommendations. He is optimistic about a three-year future in which agents can infer software needs from business goals and problems, but he also says that full autonomous buying is not here yet.

4. Learning and Application

For software companies, the first practical lesson is not simply to produce more AI-oriented content. It is to separate citation measurement from shortlist measurement. AEO dashboards, prompt panels, and synthetic prompts can help a team understand directional visibility, but the revenue-relevant question is whether the product appears in buying-intent shortlists and which trusted sources are cited around that recommendation. Optimizing only for ordinary citations may produce low-click explanatory traffic rather than influence at the evaluation moment.

The second application is to clean up machine readability. Sanders’ recommendations are operational: put an llms.txt file at the site root, block only what must be blocked for security or critical IP protection, avoid gating where possible, age out gated high-value content after 91 days, avoid hard-to-crawl PDFs, and publish in HTML or Markdown where possible. The boundary is just as important as the tactic. Crawlability is infrastructure, not reputation. A model may be able to read a vendor’s claims and still decline to recommend the vendor without third-party trust signals.

The third application is to build credible third-party evidence rather than manipulate reputation. Sanders recommends earning visibility on dependable lists, awards, and trust-signal sources, but he also says good reviews come from excellent products, not from simply asking. For product marketing teams, that changes the purpose of review programs: the goal is to increase authentic customer voice and recency, not to filter for praise. Short-term tricks may create surface visibility, but they risk damaging the trust that models, buyers, and platforms place in the evidence source.

The fourth application is to treat buyer intent as bounded account-level signal, not unlimited personal targeting. Sanders says G2 can show that people at a company such as Nike are researching a product or a competitor, but cannot identify the department or individual and cannot target specific users on a customer’s behalf. That creates a practical operating line: these signals are useful for account-level follow-up, content cadence, CRM tasks, and account-manager alerts, but they should not be treated as permission for personal surveillance.

The fifth application is to reallocate budget gradually. Sanders says companies currently spend about $60 on SEO tools, services, and SEM for every $1 spent on AEO, and he believes a 4% to 6% reallocation from search budgets in 2026 is more realistic than a dramatic shift. That percentage is his judgment, not a rule to copy blindly. A more durable approach is to fund AEO measurement, machine-readability fixes, third-party trust assets, and shortlist monitoring in small loops, then scale investment according to pipeline impact.

Finally, product leaders can use the SaaS-to-harness argument as a strategic hypothesis rather than a fixed timeline. Sanders predicts software will move from CRUD databases with business logic toward harnesses made of skills, context, governance, and connector tools, adding value on top of language models and pricing through token-consumption markup. The useful question for a roadmap is whether the product can organize models, permissions, context, connectors, and governance into reliable workflows. Adoption should still proceed by risk and trust level, especially for high-consequence actions such as purchasing, renewals, and permission changes.

Source

• Original episode: The Hidden Algorithm That Decides Which Software AI Will Recommend | Tim Sanders, G2