Website design and development

A new site starts at 3 999€ per project and covers three things: design, build and content. Design means the site gets its own shape rather than a stock template with your logo dropped on top. Build means code that loads fast and can still be extended a year from now without starting over. Content means the facts you already say out loud in a sales call end up written on the page: prices, service area, lead times, terms.

That third item is the one most proposals quietly leave with the client, and it is where projects stall. The faults we find when we measure are not design faults. They are sites where the specific number is either missing or locked inside an image, where neither a search engine nor a language model can reach it.

The build also includes the machine-facing half you never see on screen: structured data, prices as text rather than graphics, an llms.txt carrying the full offer and pricing, a robots.txt that allows AI crawlers by name, and a contact path an agent can actually use. On our own site that last item is an open lead API and an MCP server.

Source code, infrastructure and rights stay with you. One Estonian buyer wrote that condition into their tender document more clearly than most, and it is a sensible requirement at any project size, including the ones where nobody thinks to ask.

An AI answer is synthesized text, not a list of links. For a model to place you in a comparison, a readable sentence has to exist somewhere stating what you do and what it costs. A contact form contains no facts, and a price set in an image contains none either. We covered the mechanics in our guide to generative engine optimization.

We measure this with a checker that runs eight deterministic checks, with no model scoring, so the result is reproducible: whether the page opens for an agent at all, whether title and description are present, whether structured data is in place, whether robots.txt allows AI crawlers, whether llms.txt exists, whether prices are machine-readable, whether an agent can make contact, and whether an MCP interface is present.

The scores we see in the field are poor. A typical Estonian business site comes in under 50 out of 100. Our own aiside.ee scores 95 and aiasemu.ee scores 93. The faults repeat in a predictable order: prices locked in images or PDFs, no structured data, AI crawlers blocked by a privacy plugin, no llms.txt, and contact paths a machine cannot use. The index matters too: ChatGPT search and Copilot read primarily from Bing, so we submit through IndexNow after each deploy.

The honest limit: none of this guarantees an AI will recommend you. Answers vary by question and models change. The reverse is the only thing that can be guaranteed. If your facts are not readable, a model has nothing to say about you. That leaves one honest method: measure how often you are mentioned now, fix what raises the odds, measure again.

Some things do not sell in text. When someone is buying an apartment that does not exist yet, it has to be shown to them. For the EEDU apartment building we built full 3D models, a presentation film and interactive floor views where a buyer picks a floor and a unit without having to read an architectural drawing. The same project includes an operations admin, so the sales team keeps unit availability current without going back to the developer.

A second direction is calculation running inside the page. On aiasemu.ee a customer types an address and the system measures the lawn from a satellite photo and prices the job in seconds, instead of sending someone to look at it. An AI phone agent takes the calls. That is our own business, built so the methods we sell are ones we have to live with, and the numbers behind it are in our write-up of running it.

A third is mobile. Triply, our AI trip planner, runs on both iOS and Android, so an app is an available option rather than a separate world with separate suppliers.

These pieces sit in the quoted-per-project tier. The 3 999€ figure is the starting price for a complete site. 3D models, interactive floor views, an operations admin or an app go into the proposal as separate stages, so you can see what costs what and stop after the first one.

Page count is a poor measure. Ten pages repeating the same general copy are cheaper to produce than three pages that have to state concrete facts, prices and conditions. Four things move the number.

Getting the facts straight. If your price list exists and is agreed, this is quick. If prices live in three people's heads and one old spreadsheet, they have to be settled first, and that work happens whether or not it appears on an invoice.

Integrations. Whether the site has to talk to a booking system, an inventory system or a pricing calculator. Every external system brings its own access, its own failure modes and its own testing.

Languages. Two languages means twice the content work. The design is still done once.

Pages built on your own data. For aiasemu.ee we built fourteen municipality pages. The template was the same on all of them and the template was the cheap part. The data was the expensive part: cadastre records for residential land, a median yard size computed per municipality, then run through the real pricing function. The Saue page publishes 8 487 yards, a 1 135 m² median and a 65€ price; Kuusalu publishes 2 306 yards, 2 127 m² and 99€. Those pages work because they carry a number a competitor does not have.

The 990€ fix on an existing site takes about a week. A new build gets its timeline agreed at scoping, and we do not publish a single number for it, because the calendar is set by content rather than by code. The build runs in parallel with how fast the facts come together.

Five inputs keep that moving. Access to the domain, the hosting and the analytics. The price list, including a messy or partial one. Material you already hold: photos, drawings, product descriptions, previous copy. One person who answers questions and can make a decision without convening a committee. And a list of the questions customers genuinely ask, which is the single most valuable input, because those are the questions a page has to answer before an assistant has anything to quote.

We also state the result timeline up front, because it differs from the build timeline. Technical changes start to show in AI answers within weeks, since a model runs a fresh search for each answer. Mentions and reviews build over months, and no amount of technical work compresses that materially. Anyone promising a guaranteed position in AI answers within a month is selling air.

The most common honest answer we give is that you do not need a new site. If the current one looks respectable, the brand is not changing and you can still edit the content yourself, the 990€ fix is the better decision than a 3 999€ rebuild. The faults that keep a site out of AI answers are almost always technical: prices in images, no structured data, crawlers blocked by a privacy plugin, no llms.txt, a contact path a machine cannot use. None of those require a redesign, and paying for one to solve them is paying for the wrong problem. The 290€ audit settles which of the two you actually need, and its fee is credited against the work that follows. The same reasoning applies to why an assistant skips a company entirely, which we set out in why ChatGPT does not recommend your business.

Two cases where we suggest waiting altogether. First, when the offer itself is unsettled. A site that has to hide what the company does and who it does it for will not work at any budget. Second, when nobody has time to assemble the facts. Without prices, terms and service areas, a new site ends up as empty as the old one, and the budget buys nothing the old one did not already have.

One honest number from our own side. One week in PostHog on aiasemu.ee: 57 address searches, 39 refined estimates, 9 bookings started, 2 completed. Our internal conclusion was that the site is a lead machine rather than a booking machine. A website puts your facts where people and machines can both find them. Closing the deal is still your job.