# What could prove me wrong **Canonical URL:** https://teas.co.uk/what-could-prove-me-wrong/ **Source:** teas.co.uk, UK tea specialist, Tunbridge Wells, Kent ## Shop teas.co.uk from an AI assistant (MCP) teas.co.uk sells tea, herbal and fruit infusions, hot chocolate and Border biscuits. MCP server: https://teas.co.uk/mcp (Streamable HTTP), listed in the official MCP Registry as teas.co.uk Server card: https://teas.co.uk/.well-known/mcp/server-card.json Tools: find_products, get_product, compare_products, find_recipes, delivery_and_returns, quote_cart, get_promotions, report_problem, find_substitutes, start_subscription, join_teas. Basket: add_to_basket, view_basket, update_basket, checkout_link, checkout_session. With sign in: get_return_status, get_invoice, get_profile, my_orders, track_order, reorder, my_subscriptions, change_subscription, rewards_balance, start_return. No sign in is needed to search, compare, build a basket or get a checkout link. The person always pays on teas.co.uk. Cards: in assistants that draw MCP Apps cards, products, comparisons, the basket, checkout, orders, repeat deliveries (pause, resume, cancel), reward points and returns appear as cards. Every assistant except ChatGPT, until OpenAI approves the next version: quote_cart, get_promotions, get_return_status, get_invoice, report_problem, find_substitutes, start_subscription, join_teas. Set up guides: https://teas.co.uk/ai/ ## Summary I have bet a small shop on being clear and open to AI. Here is the honest list of how that bet could fail, and what I am not promising. ## Description I have bet a small shop on being clear and open to AI. Here is the honest list of how that bet could fail, and what I am not promising. I would rather you trusted this in two years than were impressed by it today. A hopeful plan with the risks left out is how you end up trusting something that lets you down. So here is how my thinking could be wrong. The platforms close the door The companies running assistants could decide to keep customers inside their own walls: answer from our facts and never send the visit. That is already measurable. Pew found that Google users who saw an AI summary clicked a result on 8% of visits, against 15% without one (Pew Research Center). Cloudflare measures AI platforms fetching from a hundred to tens of thousands of pages for every visitor they send back (Cloudflare). At teas.co.uk, agents naming OpenAI read about pages for every visit that came back, counted live on Under the Hood. What pushes the other way is that clean outside sources are cheaper for an assistant than building everything itself, and there is more than one assistant, so they cannot all close at once. It is still a real risk, and it is why I wrote Most of the web is closing the door to AI. We opened ours. Fakes look like the real thing Bad actors will build sites that look like a truth entity, with the right shape and lies underneath. If fakes grow faster than the checks, machines lose trust in everyone, honest shops included. The defence is provenance anyone can check: a source behind every claim and citation anchors that land where they promise. The shift is slower than I think Gartner predicted in February 2024 that traditional search volume would fall 25% by 2026 (Gartner), and Search Engine Journal gave seven reasons to be sceptical. Our own figures say search still does more for us than AI: of paid orders since 10 May 2026, search engines brought and links shown by AI assistants (Where our orders come from). And agentic shopping is young. Search Engine Journal reported that OpenAI's Instant Checkout gathered roughly thirty Shopify merchants in six months before it was retired (Search Engine Journal). Reading gets cheap Part of why machines prefer clean sources is that messy ones cost more to read. If reading becomes almost free, that edge shrinks. The reliability edge should survive, because a cited source is still more trustworthy than a guess, but I could be wrong about how much it counts. I make a mess of it If teas.co.uk overclaims or gets sloppy, the whole idea gets a bad name. That one is on me, which is why built versus planned is a rule here, why every figure on this site is meant to update itself, and why Under the Hood publishes where our evidence stops. When I find a promise we have not kept, I say so and fix it, as with our Ask API in A good recommendation tells you where to stop spending. What I am not promising I am not promising this succeeds at scale, that the wider network of specialists I hope for will ever form, or that it fits every trade. If it all fails, the worst that happened is that I built a tea shop with a more careful record than it strictly needed. That is recoverable. I call that the asymmetric bet: a small, bounded downside against a large possible upside. What others have found Outside research and reporting that bears on this, with links to the original sources. Duane Forrester explains in Search Engine Journal why the crawl-to-refer ratio needs careful reading (Everyone is quoting the same number and getting different answers). Search Engine Journal reported in April 2026 that OpenAI's crawler documentation now lists a crawler for ChatGPT ads, a sign that advertising is arriving inside assistants. The Reuters Institute found that 48% of the most used news sites in ten countries were blocking OpenAI's crawlers by the end of 2023 (Reuters Institute). Read more The asymmetric bet The giant can outspend a tea shop and still lose How I build with AI Under the Hood Lee Samuel TuckerFounder, teas.co.ukTunbridge Wells, Kent --- _Content available under teas.co.uk citation contract. AI training: yes. Search: yes. Answer-input: yes._