A step by step build of an AI search interface

Building an AI search interface

LGAi UnConference Workshop 1 — AI Search interface

This experiment was part of the LGAi Unconference 2026. I ran a session looking at what an AI search could look like. It's the front end, citizen facing page that we're looking at here, not what happens behind the scenes in the code. 

Each screen here is a mockup of search followed by a results page, to look at, and think about.

Some of the results are exaggerated to help the discussion. 

Some pages have a Discuss section, and you can do that in your head, with a colleague or 2, or by submitting your answers. 

‘Meridian’ is a generic search engine. Riverside council could be your council.

 

the scenario

A parent or carer in Riverside, searching for one thing: How do I apply for an EHCP for my child?

Start the walkthrough

An AI summary of the whole thing

A parent in Riverside wants to know how to apply for an EHCP for their child. A normal search engine puts three adverts, a national guide and a forum thread above the council’s own page. Riverside’s own search finds only Riverside’s pages, but ranks a 2019 PDF and an annual report among them. A search engine with AI answers confidently, with no sources, and gets the one fact that matters wrong.

Riverside’s own AI search does something narrower and more useful: it answers only from what the council has actually published, at a reading age of 9, and shows the two pages it drew on so the answer can be checked. When the question is harder, and mixes housing with an EHCP, it hands over to a person rather than guessing. It checks it has understood before it answers.

Then the walkthrough works through what stands between a parent and that answer. Nobody should have to sign in, so it remembers recent questions on the device, offers an incognito switch, and shows what signing in adds: the status of the thing you already sent. When typing is the barrier, it listens. When English isn’t the language you think in, it answers in yours, says the translation was made by a machine seconds ago, and keeps the English sources underneath to check. When reading is the barrier, it gives the same answer in Easy Read, one idea per line with a picture beside each, and says that hasn’t been checked with people who use Easy Read either.

If something isn’t on the council’s site, it says so and offers a way to ask for it to be added. With every feature switched on, the answer itself is no better than it was on step eight: sourced, and no more certain than it should be. That is the point. Not a model, but a dozen small decisions about what an AI search should refuse to do.

Written by an AI (Claude), 25 September 2026, from the walkthrough’s own pages.

Skip to a step
  1. Just a normal browser search engine, no AI Search model 01
  2. Three ads, a forum, and someone else’s council Search result 01
  3. A similar search box, on Riverside’s own site Search model 02
  4. Ten links, ranked by relevance Search result 02
  5. A search engine with AI Search model 03
  6. AI always sounds confident, even when it's wrong Search result 03
  7. A blank AI prompt, no sign in Search model 04
  8. A question, and an answer Search result 04
  9. A prompt check Search model 05
  10. A more difficult test Search result 05
  11. Remembering, without an account Search model 06
  12. Incognito, when you want no history Search model 07
  13. Or, if you’d rather sign in Search model 08
  14. When typing itself is the barrier Search model 09
  15. When English isn’t the language you think in Search model 10
  16. Same answer, their language Search result 10
  17. Ask for it in Easy Read Search model 11
  18. The same answer, in Easy Read Search result 11
  19. If it’s not there, it’s not there Search model 12
  20. Every feature, one search Search model 13
  21. More features, same answer Search result 13
  22. Not a model — the point is to make decisions for your own local authority AI search. Postscript