Last week, Google made two announcements that, taken together, reveal a dangerous imbalance in the agentic commerce race.

On July 22, Alphabet reported in its Q2 2026 earnings that Gemini now has 950 million monthly users, up from 750 million just five months earlier. Revenue hit $119.8 billion for the quarter, a 24 percent year-over-year increase. The next day, Google expanded access to Gemini Spark, its always-on AI agent, from the $99/month Ultra tier down to Google AI Pro subscribers in the United States. Spark is no longer an experiment for power users. It is a mass-market product.

Also on July 23, Axios reported that ChatGPT started integrating Yelp reviews, photos, and business information into its responses for local recommendations. Yelp is also adding a “Request a Quote” feature that lets users contact service providers directly through ChatGPT.

These developments share an underlying assumption: that AI agents should consume data from across the internet and your personal life, synthesize it, and produce recommendations you will act on. The capability is impressive. The data hygiene is not. The same agents that can infer your toddler’s nap schedule from your email history cannot reliably distinguish a real product review from a fake one.

Spark Knows Your Dog’s Name. That’s the Problem.

The Verge’s David Pierce published a remarkable hands-on with Gemini Spark on July 18. He asked Spark to plan a family weekend trip to Hershey, Pennsylvania. Within minutes, Spark produced a detailed itinerary that included:

  • Driving directions from his home (which he never provided)
  • His dog’s name (Frida, apparently found through vet emails)
  • His infant son’s age, correctly noting Lewis gets into Hershey Park free
  • His three-year-old son Arthur’s ticket requirement
  • His wife’s name and her dislike of onions and scallions
  • A Thomas Rhett and Niall Horan concert on Saturday night, pulled from Ticketmaster confirmations
  • A recommendation to get a babysitter that night

“I never told Google my dog’s name is Frida,” Pierce wrote. “My only guess is that Spark found it through emails from my vet.”

Jay Peters at The Verge had a similar experience. He asked Spark to draft an email to his wife about grocery spending. Spark found his wife’s email address (not contained in the address itself), located a budget spreadsheet in Drive (without “budget” in the filename), calculated the average, and drafted an email with a personal sign-off the couple uses privately.

This is the personalization ceiling. Google has so much data on its users that Spark can produce outputs that feel like they came from a human assistant who has known you for years. The agent is deeply, almost intimately, personalized.

Now contrast that with what happens when these same agents evaluate products.

The Product Trust Gap

When Spark, ChatGPT, or any other agent recommends a product, it consults the same data any shopper would see on Amazon: star ratings, review counts, search rankings, and sponsored placements. There is no independent verification layer. The agent processes marketplace data at face value.

This is a catastrophic asymmetry. On the personalization side, these agents are drawing from verified, first-party data: your emails, your calendar, your purchase history, your contacts. The data is accurate because you generated it. On the product recommendation side, they are drawing from a marketplace where an estimated 30 to 40 percent of reviews are fake, incentivized, or manipulated, according to multiple consumer advocacy studies. The data is inaccurate because sellers have a financial incentive to make it inaccurate.

Consider what happens when Spark plans your Hershey trip and recommends you buy a specific travel stroller from Amazon for the trip. Spark pulls the recommendation from Amazon’s search results. The top result has 4.8 stars and 2,300 reviews. Spark presents this as a strong recommendation. What Spark does not know:

  • 600 of those reviews were posted in a two-week burst following a rebate campaign
  • The seller raised the price from $89 to $129 three weeks before a Prime Day promotion, then “discounted” it to $99
  • The product ranks first because the seller bid aggressively on Amazon’s sponsored placement system, not because organic ranking put it there
  • A competing stroller with fewer but more authentic reviews scores higher on durability and safety

The agent that correctly identified your dog’s name from a vet email is now recommending a stroller based on data that is approximately 40 percent fabricated. The user, impressed by the personalization, trusts the recommendation. Why wouldn’t they? The agent just demonstrated intimate knowledge of their family. That trust transfers to the product recommendation even though the underlying data quality is completely different.

ChatGPT and Yelp: Importing Trust Problems

ChatGPT’s integration with Yelp reviews creates a parallel problem. Yelp’s review system has its own well-documented manipulation issues. Businesses routinely offer discounts in exchange for positive reviews, plant negative reviews on competitors, and game Yelp’s recommendation algorithm. A 2023 Harvard Business School study estimated that 16 percent of Yelp restaurant reviews are fraudulent.

When ChatGPT surfaces a Yelp-based recommendation, it inherits Yelp’s data quality issues without applying any additional scrutiny. The model reads the reviews, extracts sentiment, and produces a confident recommendation. The confidence comes from the model’s reasoning ability, not from the quality of the underlying data.

This is the core trust problem in agentic commerce: AI models are excellent at reasoning over data and terrible at evaluating whether the data itself is trustworthy. The more capable the model, the more convincing the output, regardless of input quality. GPT-5.6 can write a flawless analysis of why a particular restaurant is the best choice for your anniversary dinner, citing review text that was purchased for $5 on a Facebook review-seller group.

The FTC Is Moving. The Agents Are Not.

The FTC continues to crack down on deceptive marketing practices. On July 22, the commission announced it was sending $672,000 in refunds to 9,419 consumers harmed by Trend Deploy’s deceptive marketing. On July 15, the FTC finalized an order against TruHeight for deceptive supplement advertising, requiring $750,000 in penalties. These actions build on the FTC’s 2024 fake reviews rule, which gave the commission statutory authority to pursue civil penalties against companies that buy, sell, or facilitate fake reviews.

Regulatory enforcement is necessary but fundamentally reactive. The FTC punishes bad actors after the harm occurs. By the time a case reaches a settlement, thousands of consumers have already been deceived, and the fake review infrastructure has migrated to a new seller account.

AI agents need proactive protection. They need to consult data sources that have already filtered out manipulated reviews before the agent ever sees them. This is not a model capability problem. No amount of fine-tuning will teach GPT-5.6 to detect review patterns that dedicated fraud detection systems miss. It is an infrastructure problem.

What Agentic Commerce Actually Needs

The agents that consumers will trust in 2027 are not the ones with the most personal data or the smartest reasoning. They are the ones connected to the best verification infrastructure. Here is what that infrastructure looks like:

Independent review authentication. Before an agent reads a product’s reviews, those reviews should pass through an authentication layer that removes suspected fakes, detects incentivized review patterns, and weights remaining reviews by reviewer credibility. This layer must be independent from the marketplace. Amazon has no incentive to aggressively remove reviews that boost product sales. A third party does.

Price history verification. Before an agent presents a “discount” to a user, it should compare the current price against a 90-day transaction average, not against the seller-controlled reference price. This requires access to independent price tracking data. The agent should be able to say: “This product is currently listed at a 50 percent discount, but the 90-day average price suggests the real discount is 8 percent.”

Quality-adjusted ranking. Before an agent recommends a product, it should consult a ranking that reflects genuine product quality, not advertising spend. Products should be scored on review authenticity, durability signals, return rates, and competitive comparison. The agent should present seven genuinely excellent options, not the top seven results from an advertising-weighted search.

MCP-based verification calls. The Model Context Protocol, now used by Google Spark (connected to Canva, OpenTable, Instacart), Anthropic’s Claude, and dozens of other agent platforms, provides the technical plumbing for verification. Agents can call external MCP servers as easily as calling any API. A product trust MCP server that returns a Smart Score, review authenticity assessment, and price history should be a standard step in every shopping agent’s workflow.

GoBuy: The Verification Layer

GoBuy’s MCP server, available at gobuy.ai/api/mcp, provides exactly this verification infrastructure for AI agents. When an agent consults GoBuy before recommending a product, it receives:

  • A Smart Score (0-100) calculated from review quality, not quantity
  • Fake review filtering that removes suspected manipulated reviews before scoring
  • Authentic review weighting by reviewer credibility, purchase verification, and review depth
  • Top 7 product rankings per category based on genuine quality
  • GoBuy Verified status for products maintaining a Smart Score of 80+ over 90 days

An agent connected to GoBuy can tell its user: “This travel stroller has a Smart Score of 71. Eighteen percent of its reviews show patterns consistent with incentivization. It ranks fourth in its category by quality. The top-ranked alternative has a Smart Score of 88 and costs $15 less.” That is a fundamentally different recommendation than what today’s agents produce.

The difference is not the model. The same Gemini or GPT model produces both recommendations. The difference is the data layer. Connect the agent to marketplace data and you get marketplace-quality recommendations. Connect the agent to verified data and you get recommendations worth following.

The Trust Asymmetry Will Get Worse Before It Gets Better

The personalization capabilities of AI agents are accelerating. Google’s Q2 earnings show the company investing heavily in Personal Intelligence, the feature that lets Gemini reason across your Gmail, Photos, Search history, and YouTube activity. Spark’s expansion to Pro subscribers puts always-on agents in the hands of millions more users. ChatGPT’s Yelp integration is a stepping stone to broader commerce integration.

Every one of these developments makes agents more personally compelling and more commercially dangerous. The agent that knows your family intimately will earn your trust. That trust will be spent on product recommendations built on corrupted data unless verification layers are in the loop.

The companies building these agents have a choice. They can continue optimizing for personalization and capability, leaving product trust as an unsolved problem. Or they can integrate independent verification through MCP and build agents that are trustworthy end to end: personal data on one side, verified product data on the other.

Google, OpenAI, Amazon, and Meta have solved the first half. Nobody has solved the second half. GoBuy is building it.

Connect your shopping agent to verified product intelligence at gobuy.ai/api/mcp. Integration docs and examples at gobuy.ai/agent-docs.

Sources cited: