By the start of September 2026, Dinners With Friends had welcomed more than 500 guests across 50 dinners in South West England. Dates routinely sold out six to eight weeks ahead, and nearly all of the remaining 2026 calendar was booked. In November 2024, the relaunch was a napkin sketch.
Pamela Silkstone had first run the concept twenty years earlier: bring people who have never met to one table, give them a host and make the first conversation easier. I took the revived idea through its first version and first paying customers, then built the brand, website, booking system, internal tools, mobile app and search programme as demand outgrew the initial setup.
The search programme alone delivered 5,530 Google clicks in June–August 2026, up 66.1% on the previous three months, from a library of local guides that people across Exeter and Devon find while looking for food, activities and places to meet.
The live product answers the questions before asking for payment: what the dinner is, who it is for, where it happens, which dates have room and what previous guests thought. Somebody can go from the 
Dinners With Friends website to a paid reservation without leaving the service.
Get the first dinners on sale
I built the first version in Webflow with Eventbrite ticketing because Pamela needed to edit pages, change content, publish dinners and manage the site herself. It brought in the first paying customers and showed that people would pay to sit with strangers.
But the website-to-Eventbrite journey was disjointed, the feature set was basic and useful changes took too long. Pamela could run the content; I could not improve the product at the pace the dinners needed.
Replace the parts slowing the product down
In April 2025, I started a custom build with React and Next.js on the web, a Hono API and PostgreSQL underneath. Within ten days, it had a new landing experience, live event data, Eventbrite checkout, analytics, advertising attribution and a waiting list. I kept the proven payment path while replacing the parts that slowed us down.
The dinners set the requirements. Each event needed its own capacity, ticket state, guest list and messages. Sold-out dates needed waiting lists, while cancellations needed a safe way to offer seats again. In December 2025, I rebuilt the data model around event series and individual occurrences so one dinner concept could run across many dates without confusing the inventory.
Build for full tables
The common pattern became a full table six to eight weeks before the dinner, often with a waiting list. More demand meant more dates, payments and guest changes, with more chances to promise the same seat twice.
I brought booking in-house in stages: ticket tiers, a cart, temporary seat holds, a credit ledger, Stripe webhooks, refunds and disputes. The current checkout uses Stripe's Payment Element and supports guest payment. Someone can enter an email and pay; the account is created behind the transaction so the booking, attribution and follow-up stay on one customer record.
The waiting list became a workflow tied to a specific date. When a place opens, the team can send a timed offer, check whether it was delivered, resend it or revoke it. A hold protects the seat while the guest decides.
Run the whole dinner operation
Customers can discover dinners, pay, manage reservations, leave reviews, make referrals, receive notifications and use an inbox. Internally, the team can set up dinners, create dates in bulk, manage ticket tiers, search attendees, transfer bookings, check guests in, offer waitlist places and communicate through inbox, push and email.
Venues change, places reopen, reminders come due and hosts need answers before service. I built previews, scheduling, delivery records and audience rules into the messaging tools so the team can see what was sent and what happened next.
I also built an Expo app for customers and hosts, covering event discovery, bookings, reservations, waitlists, credits, maps, conversation prompts, check-in and push notifications. Android reached an installable release build and the Play internal track. iOS reached launch preparation with runtime configuration, privacy metadata, review notes and store screenshots.
The visual system had to work for warm, social dinners and denser account, payment and administration screens. I iterated the brand before settling on the current aubergine, cream and green system across the website, checkout, app and campaign work.
Move the live service without losing customers
By July 2026, the original Supabase setup was getting in the way of how I wanted to run the product. I moved the live service to PostgreSQL, Better Auth, Redis and object storage on
Railway.
I rehearsed the migration against a fresh copy of live data before cutting over. The production move preserved user IDs and password hashes, transferred stored media, kept the API contract stable and left every public URL in place. Existing customers kept their passwords, and local guides kept the addresses Google and readers already knew.
That gave me control over data, authentication, jobs and storage without interrupting people using the service.
Grow Google search clicks by 66% in three months
I built an MDX publication with page metadata, open-graph images, event and venue schema, breadcrumbs, internal links, sitemaps and IndexNow. The working loop was to publish something useful, measure it and improve the pages closest to working.
In April 2026, I turned that base into a maintained local guide library. I deeply rewrote seven leading guides, then researched and updated internal links across another 115 articles. Search Console data became a work queue: keep dated snapshots, find pages within reach of page one, estimate clicks lost to a weak title or description and review the highest-value opportunity first.
The content inventory joins articles on disk to pages Google sees, counts named venues likely to go stale and weights that risk by search exposure. Audits found old businesses and poorly supported claims, so I added closed-venue registries, recurring-claim checks and primary-source review.
Search Console recorded 9,131 clicks and 640,956 impressions between 29 April 2025 and 28 August 2026, with a 1.4% click-through rate and an average position of 10.8. Under Google's definition, those figures measure appearances and clicks in search results, not bookings or revenue.
From June to August 2026, Google web search delivered 5,530 clicks, compared with 3,330 in March–May: 2,200 additional clicks, up 66.1%. Impressions rose from 266,718 to 381,345, up 43%. These were complete, consecutive 92-day periods, checked in Search Console on 6 September 2026. Average CTR rose from 1.2% to 1.5%, while average position moved from 9.8 to 11.3. Search traffic rose, but this comparison does not isolate seasonality or attribute dinner bookings to SEO.
The homepage led the landing-page table, followed by guides to afternoon tea, pub quizzes, restaurants, fish and chips, running clubs and ice cream. People found the site while making a local decision, got a useful answer and could meet the dinner product on the same visit.
Between March–May and June–August 2026, the refreshed pub quiz guide grew from 76 to 226 clicks, the running clubs guide from 70 to 199, and the fish and chips guide from 86 to 188. Each more than doubled its Google clicks. The guides give the business entry points beyond searches for its own name.
Separate AI fetches, citations and referrals
I measure AI discovery in three layers: a crawler fetch means an engine accessed a page; a citation means the page appeared as a source in an answer; a referred session means a person clicked through. Combining them as “AI traffic” would obscure what happened.
Middleware recognises assistant and crawler user agents and records an ai_bot_fetch event with the bot and requested path. A separate browser tracker records referrals from ChatGPT, Perplexity, Claude, Gemini, Copilot and Duck.ai. Google Search Console measures generative-search exposure; Bing measures citations and grounding queries across Microsoft Copilots and partners.
PostHog counted 24,376 ai_bot_fetch events over the last 90 days, including ChatGPT-User, OAI-SearchBot, ClaudeBot, PerplexityBot and Bytespider. Those are machine fetches, not 24,376 people, visits or bookings.
Google Search Console's beta Generative AI report showed 30.7K impressions in the available data. The leading pages were guides to Exeter neighbourhoods, Dawlish and Teignmouth, a day out in Sidmouth, beaches near Exeter and the psychology of sharing food. Google exposes impressions there, not clicks or citations.
Bing Webmaster Tools reported 2.8K citations between 30 May and 29 August 2026, with an average 11 unique Dinners With Friends pages cited per day. Microsoft describes this report as aggregated data from Copilot, Bing AI answers and selected partners, and says its grounding queries are sampled. The homepage led, but local, friendship and food guides were cited too.
What the full tables required
I didn't start by designing a large platform. I started with a Webflow site and Eventbrite tickets, took the first payments and built what the filled dinners demanded. By September 2026, nearly all remaining dates were booked. Behind those tables is a product I can keep changing: owned payments, bookings and customer data; an app for guests and hosts; infrastructure I control; and search and AI-discovery reporting that separates attention from customers.
WithSeismicWork with meI build and run businesses with a team of me plus AI agents, and help you do the same.
