The middleman tax is dying — and AI shopping agents are the cause

For 15 years, every small business decision has been shaped by gatekeepers. Run a restaurant in Tokyo? You list on Tabelog or Hot Pepper Gourmet. A salon? Hot Pepper Beauty or StyleSeat. A hotel? Booking.com or Expedia. A takeout shop? Uber Eats or DoorDash. The deal was always the same: the platform brings you customers, you give up a margin. 15%. 25%. 30%. Forever.

The platforms got rich. Customers got convenience. Businesses got squeezed. Booking Holdings reported $26.9 billion in 2025 revenue, almost entirely from commissions. Combined US restaurant commissions through DoorDash and Uber Eats alone exceed $15 billion per year. The structure of platform aggregation has been the dominant force shaping SMB economics since 2010.

That structure is starting to crack. Not because the platforms are weakening, but because customers are starting to skip them entirely — by asking AI assistants instead.

The middleman tax, by the numbers

The fees most SMBs pay are not subtle. Here's what came back from a verification pass against each platform's published rate cards:

Restaurants and food delivery. Uber Eats charges 15-30% per delivery order depending on tier (15% Lite, 25% Plus, 30% Premium). DoorDash charges the same 15-30% range (15% Basic, 25% Plus, 30% Premier). OpenTable charges $1 per cover via OpenTable.com plus monthly subscription fees ranging from $39 to $449. In Japan, Tabelog and Hot Pepper Gourmet's specific rates are not publicly disclosed — owners receive private quotes — with industry trade press citing premium plans in the ¥120,000 to ¥1.2 million per year range per restaurant.

Hotels and travel. Booking.com's industry-standard 15% commission rises to 17-25% in major cities and through Preferred Partner tiers. Expedia operates in the same band, with rate cards varying by Expedia Traveler Preference program tier. Airbnb splits its fee: roughly 3% from the host plus 14-16.5% from the guest for most listings, or a single ~15.5% host-only fee mandatory for hotels and PMS users. Cloudbeds reports the current OTA average is 15-30%+.

Beauty, salons, and wellness. Hot Pepper Beauty's listing fees + per-cover commissions are not publicly disclosed by Recruit; trade press estimates ¥30,000 to ¥500,000 per salon per month. StyleSeat charges $35 per month flat. Booksy charges $29.99 per month plus 30% of new client first visits acquired through their Boost feature. Mindbody starts at $99 per location per month on the Starter tier.

Online marketplaces. Etsy charges 6.5% per transaction plus a $0.20 listing fee plus 3% + $0.25 payment processing, and 12-15% offsite ads become mandatory once a seller crosses $10,000 in annual revenue. Amazon charges 8-17% referral fees depending on category (most categories ~15%) plus a $39.99 monthly Pro subscription. Shopify subscription tiers run $39 to $2,300+ per month plus 2.9-3.55% payment processing — and a 0.6-2% non-Shopify-Payments fee on top.

Add it up. A typical $100,000-revenue SMB pays roughly $15,000 to $25,000 per year to platform intermediaries before any cost of fulfillment. A bottom-up triangulation suggests US-only platforms in this list extract $150-200 billion+ per year from small businesses. Japan's Kakaku.com, Recruit, and Demae-can add another ¥500 billion+. No single analyst report aggregates this — partly because, as the Tabelog and Hot Pepper opacity shows, the middleman tax is partly hidden by design.

What changed in 2025

From an SMB's perspective, the platforms looked unbeatable for fifteen years. Then the AI agentic shopping infrastructure landed in a single 18-month window:

• October 2024 — Anthropic launched Claude Computer Use, giving an AI the ability to drive a browser like a human. Early users included DoorDash, Replit, and Asana.
• November 2024 — Perplexity launched its Shopping Hub, backed by Amazon and Nvidia, letting users buy products directly inside Perplexity.
• January 2025 — OpenAI launched Operator at $200/month for Pro users, with launch partners DoorDash, eBay, Instacart, Priceline, StubHub, and Uber.
• April 2025 — Microsoft Copilot Shopping Actions went live, integrating Expedia, OpenTable, Instacart, Kayak, Klarna, and the Shop-from-Shopify network.
• September 2025 — OpenAI and Stripe shipped the Agentic Commerce Protocol, enabling Buy-in-ChatGPT for native checkout. Etsy was the first integrated merchant.
• December 2025 — Stripe released its Agentic Commerce Suite, the merchant-side infrastructure for AI-driven checkout.
• January 2026 — Google announced Universal Commerce Protocol (UCP) with launch partners Walmart, Home Depot, Wayfair, and Urban Outfitters, plus native shopping integration in Gemini.

And the consumer-side data, where it's verifiable, is striking. Alipay AI Pay clears more than 120 million weekly agent-mediated transactions in China today — the only large, real consumer-scale number anywhere. McKinsey projects up to $1 trillion of US B2C retail revenue could be orchestrated through agentic commerce by 2030, with $3-5 trillion globally.

There's also the honest counter-weight: the Wall Street Journal reported in November 2025 that few companies deploying AI agents have received a return on investment, and Gartner placed generative AI in its "trough of disillusionment." The infrastructure is finished. The behavior shift has started. Mass-market consumer ROI is not yet proven.

The DTC parallel — and what it teaches us

If this all sounds familiar, it should. We've seen this movie before. From 2010 to 2018, a wave of direct-to-consumer brands attempted exactly the same disintermediation thesis: skip the retail middleman, sell direct to the consumer, and capture the margin the gatekeeper used to keep.

Warby Parker bypassed Luxottica's eyewear monopoly. The economics were stark: LensCrafters founder E. Dean Butler told the Los Angeles Times in 2019 that quality frames cost "$4 to $8" wholesale and designer-quality frames "around $15," calling retail prices "a complete rip-off." Warby launched glasses at $95 against Luxottica's $300-500, reclaiming roughly the 85% margin that vertical integration plus skipping the licensee chain produced.

Dollar Shave Club pulled the same trick on Gillette. Casper on Tempur Sealy. Glossier on Sephora and Estée Lauder. Allbirds on Nike. Harry's on the same razor giants Dollar Shave Club went after. They all worked, briefly. Dollar Shave Club sold to Unilever for ~$1 billion in July 2016. Warby Parker direct-listed on the NYSE in September 2021.

Then they all stalled. Casper IPO'd in February 2020 at a 57% haircut from its private valuation, and was acquired by mattress conglomerate Carpenter Co. in October 2024. Allbirds went from a $4 billion+ peak market cap to $21 million in April 2026, before being sold to American Exchange Group for $39 million. Glossier, once valued at $1.2 billion, closed 9 of 12 stores by 2026 and started wholesaling through Sephora — the very middleman it was built to skip.

What killed DTC? Three forces. First, the Facebook tax. DTC scaled cheaply on Meta ads from 2012 through 2018. Then Apple's App Tracking Transparency in iOS 14.5 (April 2021) hit, with Flurry Analytics reporting 96% of US users opted out of IDFA sharing. Customer acquisition costs spiked overnight. Meta itself disclosed roughly $10 billion in revenue impact. Second, IPO reality check — public markets stopped paying DTC valuations. Third, hybrid retail won — every survivor pivoted off pure DTC into Sephora, Target, Nordstrom, or built physical stores.

The lesson is brutal and useful: DTC didn't disintermediate distribution — it traded one middleman for another. Replacing LensCrafters with Facebook Ads + Shopify + UPS meant Meta, Shopify, and the carriers extracted what Luxottica used to. When Apple changed the rules unilaterally in 2021, the entire cost base of every DTC brand inflated at once.

Why AI agents might be different

Here's where the AI shopping shift could break the pattern: structural fragmentation. Unlike 2014's Facebook plus Google duopoly, the 2026 AI shopping layer has at least seven serious competitors — ChatGPT, Perplexity, Claude, Gemini, Microsoft Copilot, Apple Intelligence (default-on-device on every iPhone), and Amazon Rufus. None can extract Facebook-style rents because users — and the businesses being recommended — can route around any one of them.

Multi-agent pluralism is the closest the internet has come to genuine disintermediation since email. Email is the only durable example of a discovery channel where no single company captures most of the rent. If AI shopping stays multi-platform, the new "shelf" inside ChatGPT or Claude or Perplexity has no toll booth, because the toll booth keeps moving.

But — and this is the honest counter-thesis — every prior "disintermediation" wave produced a new aggregator. Amazon replaced shopping malls. Google replaced the Yellow Pages. Booking.com replaced travel agents. Uber replaced dispatch. Aggregation is gravity. If one AI agent wins category share — plausibly OpenAI given ChatGPT's lead, or Apple via default placement on every iPhone — the new tax becomes agent-placement fees, not the absence of a middleman. That's already happening at the edges: OpenAI and Stripe's Buy-in-ChatGPT charges merchants an undisclosed cut on Etsy transactions. Google's UCP first partners are Walmart, Home Depot, Wayfair, and Urban Outfitters — the agent-shopping version of Amazon's marketplace.

The most likely 2030 outcome is not "agents replaced the middlemen" but "agents became the new SEO" — a thin layer where incumbents pay to rank, with margins compressed but not eliminated. The DTC parallel suggests a 10-year window of arbitrage followed by re-aggregation, not permanent disintermediation.

But for an SMB looking at the next 3-5 years, the arbitrage window is exactly what matters.

What an SMB should do this week

Three groups of actions, ranked by leverage. We recommend doing the first group yourself. The second group is harder; we built our $47 audit to score it for you. The third group is the build work most owners hire out.

Group 1: DIY this week

• Audit your robots.txt. Add explicit `Allow: /` rules for `OAI-SearchBot`, `Claude-SearchBot`, and `PerplexityBot`. These are the bots that surface your business in answers — different from training bots like `GPTBot` and `ClaudeBot`, which you can choose to block separately. OpenAI documents the bot list here, Anthropic here, and Perplexity here.
• Claim your Google Business Profile. Free, takes an hour at business.google.com. This is the single biggest entity signal AI engines triangulate against, and it costs nothing.
• Run the discoverability test. Open ChatGPT, Claude, and Perplexity. Ask each: "best [your category] in [your city]." Three times. Fresh sessions. Record whether you appear and who's cited. This is your baseline.
• Add llms.txt at the root following the Howard 2024 spec at llmstxt.org. Cheap insurance. Some agentic tools and developer clients already consume it; no AI lab has officially confirmed reading it for citation, but it's a 5-minute setup.

Group 2: Score before you fix

The next set of fixes — Schema.org structured data, content rewrites using the GEO-paper levers, entity disambiguation across Wikidata + LinkedIn + GBP — is harder to DIY because the value comes from doing them correctly, not just doing them. A misconfigured `LocalBusiness` schema is invisible. A schema with wrong subtype or missing `geo` coordinates is worse than no schema. This is the work our $47 audit scores: where you stand against the 30-item LLM-friendly checklist, ranked by impact.

Group 3: Build with structure

Once you know what's missing, three categories of fixes deliver outsized AI-citation lift:

Schema.org coverage. Schema.org/LocalBusiness plus the most specific subtype (Restaurant, DaySpa, Dentist, etc.), with `name`, `address` (full PostalAddress), `geo` at 5+ decimal places, `openingHoursSpecification`, `telephone`, `priceRange`, and `image`. Add `Organization` for the legal entity, `Service` nodes for each offering, and `sameAs` linking to your Wikidata Q-item, Google Business Profile, LinkedIn, and Facebook. Validate at search.google.com/test/rich-results.

Content patterns that lift AI citation. The KDD 2024 GEO paper (arXiv 2311.09735) by Aggarwal et al. tested seven content rewrites against Perplexity.ai. Three lift AI citation by 30-40%: direct quotations from named sources (+41% — the highest), statistics with cited numbers (+31% overall, +37% on Perplexity specifically), and inline citations to authoritative sources (+27% overall, +115% on currently low-ranked pages — meaning the tactic actively democratizes who gets cited). Two patterns did NOT help: keyword density, and authoritative tone alone. One actively hurt: keyword stuffing measured ~10% worse than baseline.

Server-side rendering and clean URLs. Google's own JavaScript SEO documentation confirms that not all bots can run JavaScript, and AI crawlers (`GPTBot`, `ClaudeBot`, `PerplexityBot`) do not document JS execution. Assume they read raw HTML only. Server-side rendering or pre-rendering is non-negotiable for AI visibility.

The bet

AI shopping agents are not a magic wand. Most consumers in 2026 still discover businesses through Google, Instagram, and word of mouth. The shift to AI-driven discovery will take years to fully materialize — possibly decades, possibly never if AI consolidation gives us a new Google we just renamed.

But the work to be AI-friendly is also the work to be trustworthy on the open web. Server-side rendering helps Google. Schema.org markup helps Google. Citations and quotes help readers. Allowing AI bots to crawl your site costs nothing if the AI shift fails. It costs you everything if it succeeds and you're not on the shelf.

We built BizAIReady because the gap between SMBs and the AI shelf is mostly mechanical, not strategic. Most owners don't need a marketing rethink — they need their robots.txt, schema, and content patterns to match what the new agents actually read. A $47 audit shows where you stand. A one-time $497-$1,997 build fixes the gaps. After that, you own your discoverability instead of renting it. No commission, no listing fee, no monthly subscription to anyone — including us.

That's the entire pitch. The middleman tax is starting to die. The question for every small business is whether you'll still be paying it when it's gone.