What a company website actually needs, and what you are being sold instead

Three quotes, three different diagnoses, and no way to tell which one is real. What to insist on, what to ignore, and the questions to ask before you sign. Written August 2026.

Get a quote

Three agencies quote for your new website. The first has a line item called “AI search optimisation”. The second has “FAQ structured data implementation”. The third has neither, but tells you your current site “doesn’t convert”, so what you need is a redesign.

All three sound plausible, and all three use words you have no way of checking.

I will check them for you, starting with the boring part, because it matters more.

Five things a company website always needs

A company website has five jobs, and all five are the same as they were a decade ago.

It has to load fast and say what the company does in the first sentence, in the words an outsider would use. It has to work on a phone, because that is where most of your visitors are. The people who work at the company should be able to edit it themselves. And it has to survive one person leaving.

That last one kills more websites than everything else combined. Companies lose control of their own domain because the registration sat on the personal account of someone who left two years ago and does not answer his phone. The fix is one sheet of paper: who holds the domain, who holds the hosting, where the code lives, whose account the analytics sit on, who owns the ad account. Write that sheet before you sign the contract.

Most articles about websites in 2026 open with what is new. I started with these five because if they are broken, nothing new will fix them.

What AI Overviews changed, and what GEO is

One thing did change, and half the industry now makes its money from you understanding it imprecisely.

A growing share of your readers meet a summary of your site before they meet your site. Google shows AI Overviews, and some people go straight to ChatGPT or Perplexity. Senuto analysed 17,763,868 Polish keywords and found AI Overviews on 28.95% of them, roughly one query in three (Senuto report, edition 2).

Two other numbers in that report matter more to you.

First: in the earlier edition of the report, 20.68% of citations came from sources absent from the organic results, which read like a separate way in for a small company. In the second edition, on the same keyword base, that figure fell to 5.34%, and 49% of citations come from the organic top ten (edition 2). So there is no separate channel to buy: the work that earns a citation in AI is the same work that earns a position in the results. If somebody sells you GEO as a thing beside SEO, show them those two numbers.

Citations from sources absent from the organic results
  • Earlier edition of the report20.68%
  • Second edition, same keywords5.34%

Second: more than 96% of citations in Polish AI Overviews come from Polish sources, and the foreign share is lower still in the newer edition. If you run a bilingual site and one language is a translation bolted on in the final week of the project, you have handed that channel to someone else. It cuts both ways: a Polish company entering an English-speaking market with machine-translated English is making the same mistake in the other direction.

of 17,763,868 Polish keywords show AI Overviews
28.95%
of citations come from the organic top ten
49%
of citations come from Polish sources
over 96%

The practical consequence is that text a machine can parse and quote wins: definitions, numbers, comparisons, procedures, the names of things. A page that says “professional solutions for demanding clients” will never be quoted, because there is nothing in it to quote. A page that says “we install industrial doors within 80 km of our depot, three weeks is the standard lead time” will be.

There is a second consequence that almost never comes up during redesign meetings. If a model cites specific URLs, then changing every URL during a rebuild cuts you out of those summaries too, for months. That alone makes stable URLs worth protecting. A builder who plans a rebuild without a redirect map from old addresses to new ones is throwing away visibility you already paid for.

All of this is the advice good studios were already giving in 2015: write plainly, structure your content, do not change URLs without a reason. Now there is a measurable reward for following it.

Four things you will be sold that do not work

This part will annoy some people.

FAQ structured data. On 7 May 2026 Google marked its FAQ structured data documentation as deprecated, and FAQ rich results stopped appearing in search (Search Engine Journal). The process began in August 2023, when Google restricted them to government and health sites, so for your company they have not worked for nearly three years. The Search Console report went in June 2026. If the quote in your hand has a line for “FAQ schema implementation for improved visibility”, that line bills you for something Google does not display. FAQPage remains a valid Schema.org type and it harms nothing; the mistake is paying for it as a source of traffic.

llms.txt. A text file in your site root, sold as the way to be visible to AI. Ahrefs examined 137,210 domains and checked who fetches these files. 97% of them received zero requests in May 2026. Among the 3% that received anything at all, SEO audit tools accounted for 21.7% of requests and every AI bot category combined for 19.5% (Ahrefs study). So a file built for AI gets fetched more often by SEO crawlers than by AI bots.

Requests to llms.txt files, among the 3% that received any
  • SEO audit tools21.7%
  • All AI bot categories combined19.5%

Reboot went further with a controlled experiment: four pages on each of two domains, with no internal or external links pointing at them, listed only in llms.txt. After three months no AI bot had visited any of them, while the same bots were crawling the rest of both sites normally (Reboot experiment).

In fairness, ENSOMEDIA publishes an llms.txt file too, and it is probably doing nothing. It took an hour to write, and I keep it current because it doubles as a short description of what I do, which is useful to me internally. I do not sell it as a service, and I would not advise you to pay anyone for it.

The “40 to 60 words” rule. You will often hear that an answer should run 40 to 60 words to win the snippet. That number describes how many words fit inside Google’s box. Portent measured 7,854 featured snippets and said so directly: “I wouldn’t say 40-55 words is ‘optimal,’ since I doubt Google is looking for a particular word count” (Portent study). Someone measured the size of a box and the industry turned it into a recipe. Write the answer at the length the question deserves, and keep it short where short reads better.

Schema as a source of AI citations. Ahrefs tracked 1,885 pages that added JSON-LD structured data, matched them against around 4,000 control pages, and measured citations across ChatGPT, AI Mode and AI Overviews. There was no meaningful lift (Ahrefs study). The correlation is real and strong, because sites with clean schema tend to be sites run properly in every other respect. Add basic structured data, because it is cheap and it tidies up your company information, and expect no extra citations from it.

Template or a custom company website

The answer is “template” far more often than most agencies will admit.

One service, ten pages, no catalogue, no integrations: a well-chosen template on a decent CMS does the job and leaves budget for the things that move the result, which are photography and words. A company that spends everything on a unique layout and then fills it with stock images and copy paraphrased from a competitor loses to a plain template filled with the truth.

Custom starts to make sense when the site stops being a site: when it has to log people in, run two languages editorially, or talk to a warehouse, a CRM or a booking system. It also makes sense when there is so much content that the layout has to be designed around its structure.

A concrete case from my own work: dotlaw.co is a bilingual publishing platform for a law firm, built on Next.js and Sanity, with semantic search across the magazine and stable routes. No template does that, because the hard part was letting lawyers publish in Polish and English without a developer standing in the middle. Had the same firm needed five pages and a contact form, I would have told them to buy a template.

“We need a redesign” usually means something else

When somebody says the site looks dated, nine times out of ten they mean the site says nothing specific.

Test it before you spend anything. Open your homepage and read the first screen aloud to someone outside your industry. If that person cannot repeat back what you do, the problem sits in the words, and a new layout would only package it more attractively.

Page speed is mostly a content decision too. Sites are slow because somebody put nine slides in a carousel, a background video, four typefaces and six tracking scripts of which exactly one gets used. A developer can optimise that and take half a second off your load time, while deleting the carousel takes off three.

What to ask before you sign

Five questions, and each one should get a specific answer.

  1. Whose accounts will hold the domain, the hosting and the analytics, and when do I get access? The correct answer is yours, from day one.
  2. Who on my team can change the text on the homepage, and how long will it take them? If the answer is “you email us”, price that as a fixed cost for the life of the site.
  3. What happens if we stop working together in two years? Ask to be shown where the code lives and what a handover looks like.
  4. Is there a line in this quote that somebody else would call unnecessary, and what is the case for keeping it? A builder worth hiring will tell you which lines are optional; a salesperson will call all of them essential.
  5. Who is writing the words? This decides the outcome more than the choice of framework, and it usually comes up last, if it comes up at all.

There is a sixth question clients rarely ask: what happens to this website on the day you stop thinking about it? Sites usually die because nobody added a sentence to them for three years. If nothing in the quote makes a named person responsible for the site continuing to change after launch, you are buying a monument.

If you do one thing this week off the back of this article, write the access sheet. It takes twenty minutes, and it is the only item on the list that needs nobody’s approval and no invoice.

From here there are two routes, depending on where you are.

You are holding quotes and want a second reading. Describe the project in the project brief and send the lines you cannot check. I will tell you which of them I would refuse to bill for myself.

You already know what you want and mainly want to know the price. One closed scope on this site has its figure published in advance: a simple company website, in what a simple company website costs, with the whole scope, every exclusion and the hours behind the number. How I build company websites, and two that are already up, is on the websites page.

Frequently asked questions

What does a company website actually need?

Five things, all unchanged for a decade. It has to load fast, say what the company does in the first sentence in the words an outsider would use, work on a phone, be editable by the people who work there, and survive one person leaving. More websites fail on the last one than on the other four together.

Does FAQ structured data still work?

It no longer works as a source of traffic. Google marked the documentation deprecated on 7 May 2026 and FAQ rich results stopped appearing; the restriction to government and health sites began back in August 2023. FAQPage remains a valid Schema.org type and harms nothing. Paying for it as a visibility improvement means paying for something Google no longer shows.

Does llms.txt help with visibility in AI?

There is no evidence that it does. Ahrefs checked 137,210 domains and 97% of their llms.txt files received zero requests in May 2026; among the 3% that got anything, SEO audit tools accounted for more requests than every AI bot category combined. Reboot listed four unlinked pages only in llms.txt and after three months no AI bot had visited any of them.

Is GEO something separate from SEO?

On the current numbers, no. Citations from sources absent from the organic results fell from 20.68% to 5.34% between two editions of the same Senuto study, and 49% of citations now come from the organic top ten. Earning a citation in AI takes the same work as ranking in the organic results.

Ask about a company website

Say what the company does and what you need on the website. Leave your email for a reply.

Got an idea?
Let’s build it.

© ENSOMEDIA 2026 Privacy PolicyBack to top