Iron Goo
Two studio professionals check a laptop beside a quote form and phone in a warmly lit office on a rainy night.

Does UX Still Matter When AI Is the Interface?

Atamyrat Hangeldiyev
Atamyrat Hangeldiyev
Systems Architect
March 5, 2026
On this page
UX

Yes, UX still matters, and the most useful place to inspect it is where a visitor checks a fact or takes action. In a Pew study of 900 US adults browsing in March 2025, people clicked a traditional result on 8% of visits to Google results pages with an AI summary, versus 15% without one. They clicked a link inside the summary on about 1% of all visits to results pages that showed a summary. These are visit-level rates, not link click-through rates. Pew Research Center published the findings on 22 July 2025.

A visitor who reaches your site may have already read an answer and arrived to check a price, confirm a fact, or book. That is an inference about the journey, not something Pew measured for each visitor. Your booking flow still has to work for that person, and your page structure has to make sense to software acting on their behalf. The wider scope of UX for a small business website includes those moments as well as the route into them.

8% vs 15%
Visits with a traditional result click
22.43%
AI conversations followed by verification
8 square pixels
Agent guidance for interactive elements
MythWhat the source saysVerdictCheck to run
AI assistants will replace websites, so UX is deadNN/g, August 2025: no participant relied entirely on generative AI for every information need.WrongComplete your main task from its landing page.
Nobody clicks through from AI answersPew, July 2025: traditional results still received clicks on visits with summaries.WrongCheck whether a visitor can find the next action on a landing page.
Chat is the best interface for everythingJakob Nielsen, June 2023: future AI interfaces will likely combine interaction styles.Half trueCount the fields a visitor must specify to request a quote.
Agents can work around a clumsy siteGoogle, July 2026: browser agents can inspect visual renderings, the DOM and the accessibility tree.WrongInspect your main call to action in the accessibility tree.
AI-generated design makes UX skills obsoleteNN/g, October 2025: prototyping tools need extensive human guidance for design tradeoffs.WrongTest a generated screen against a real task.
Users trust AI answers, so sites need no trust signalsNN/g, September 2023: participants verified some answers, especially when stakes were high.WrongFind your price or pricing basis in page text.
If the AI summary is right, conversion followsAdobe, March 2025: AI-referred retail visitors were still less likely to convert than other visitors.WrongCompare AI Assistant and Organic Search key-event rates.

Myth 1: AI assistants will replace websites, so UX is dead

What people say

An assistant can answer a question without sending its user to a website. If that happens often enough, the argument goes, a company can stop treating its website as a place where people make decisions.

What the source shows

In an August 2025 qualitative study of nine participants, NN/g found that “Nobody relied entirely on genAI’s responses (in chat or in an AI overview) for all their information-seeking needs.” Participants often used search and AI chat in tandem to explore the same topic and sometimes used them to fact-check each other.

Google’s AI optimization guidance says Google Search’s generative AI models use publicly accessible, crawlable web content. On 22 July 2026, Sundar Pichai also said AI features were sending billions of website clicks a week, a claim made by Google rather than an independently established traffic count. A site can therefore serve both as a source of information and as the place where someone completes a task.

The practical priority is the action surface: the price or pricing basis, the booking control, the form, and the information needed to use them. A page that gives a polished introduction but makes a visitor hunt for availability has left the decision unfinished.

Check this on your site

Open your most important service landing page on a phone. Starting at the top, try to find the price or pricing basis and reach the booking or inquiry form; record each missing fact and each screen you have to pass through.

Myth 2: Nobody clicks through from AI answers

What people say

An answer at the top of a search page satisfies everyone, so any investment in the page behind it is wasted. The smaller percentage in a click study can make that conclusion look tidy.

What the source shows

Pew’s figures describe a substantial difference, not an absence of clicks. On visits to results pages with an AI summary, 8% included a click on a traditional result; the corresponding figure was 15% without a summary. A link inside the summary was clicked on about 1% of all visits to results pages that showed a summary.

Seer’s analysis found that click-through rate on queries with an AI Overview reached a 1.31% floor in December 2025 and recovered to 2.36% in February 2026. Its transactional segment where the brand was not cited fell from 4.17% to 2.15% during 2025 and returned to 4.21% in February 2026. Seer says it cannot claim causation.

The query mix matters. In Seer’s window, AI Overviews appeared on 36% of informational queries, 8% of commercial queries and 5% of transactional queries. Those figures describe that dataset, not a guaranteed exposure rate for your company. The comparison of SEO click studies examines the measurements in more detail.

Hypothetical: applying Pew visit-level rates to 10,000 visits

10,000 visits without a summary × 15% = 1,500 visits with a traditional result click 10,000 visits with a summary × 8% = 800 visits with a traditional result click 10,000 visits with a summary × about 1% = about 100 visits with a summary-link click

This arithmetic uses two hypothetical groups of 10,000 results-page visits. Pew measured behavior across queries; its rates do not forecast what will happen to one search result or one website. The summary-link figure is a separate measure, so the calculation does not assume that the two types of click came from different people.

An AI answer sends fewer visitors to a website, where a visitor reaches a booking form.

Notice the smaller group reaching the site and the task waiting for them there. A visitor who has already researched an answer may need a direct route to booking.

Check this on your site

Choose a page that receives search traffic and open it at a mobile width. Check whether its first screen identifies the service and gives a visible route to the next action without requiring a visitor to return to search.

Myth 3: Chat is the best interface for everything

What people say

If a visitor can describe what they want in a sentence, a chat box can replace navigation, forms and filters. That can sound appealing when your existing navigation has grown into a museum of old department names.

What the source shows

In his 18 June 2023 opinion essay, Jakob Nielsen described AI interaction as a move toward specifying an outcome. He also identified a problem with chat: users must express their problems in prose. His proposed direction was a mixture of interaction styles.

Future AI systems will likely have a hybrid user interface that combines elements of both intent-based and command-based interfaces while still retaining many GUI elements.

Jakob Nielsen, AI: First New UI Paradigm in 60 Years

Nielsen estimated that half the population in rich countries may not be articulate enough to get good results from current AI bots. That is his opinion based on literacy research, not a measurement from an interface test. Nielsen’s essay gives the reasoning behind it.

In an NN/g observational study published on 27 February 2026, participants used search to validate information and obtain specific details. Several said they would not fully trust AI-provided pricing. A table can make prices comparable, a date picker can expose available dates, and a filter can show the options a company actually offers.

A chat feature can still help someone describe an unusual requirement. Give it structured controls when the answer depends on a date, an option or a price, and let the visitor see what they selected. For a feature on your own site, the guide to designing AI features for a small business addresses that interaction.

Check this on your site

Open your quote request flow and count the facts a visitor must provide, such as service type, location or preferred date. Check whether each fact has a visible field or choice and whether the visitor can review it before submitting.

Myth 4: Agents can work around a clumsy site

What people say

An AI agent is intelligent, so it can infer what an unlabeled control does and find its way around an overlay. The site owner can leave the interface alone and let the agent compensate.

What the source shows

Google’s AI optimization guidance, updated 10 July 2026, describes several ways a browser agent may examine a page:

AI agents are autonomous systems that can perform tasks on behalf of people, such as booking a reservation or comparing product specifications. These agents can take many forms; for example, browser agents may access your website to gather the data they need to complete these tasks, such as analyzing visual renderings (like screenshots), inspecting the DOM structure, and interpreting the accessibility tree.

Google Search Central, AI optimization guide

The Google guidance does not say every agent uses every view. It does give you a useful inspection rule: a control should make sense visually and expose a useful role and name in the page structure.

A browser agent reads a booking button through a screenshot, the DOM and the accessibility tree.

The same page appears as a visual rendering, a DOM structure and an accessibility tree. Notice how the booking button needs an understandable action in each view.

OpenAI’s January 2025 Operator launch described a browser-based system that looked at pages and interacted by clicking, typing and scrolling. Its July 2025 ChatGPT agent launch described a visual browser alongside a text-based browser. These are past product launches, and their interfaces can change; the underlying site task remains recognizable.

In their web.dev guide, updated 1 April 2026, Kasper Kulikowski and Omkar More recommend avoiding transparent overlays that cover controls and keeping layouts stable. They also recommend native button and a elements, or an appropriate role and tabindex for modified elements; associating labels with inputs; using cursor: pointer; and making interactive elements needed for an action visible and larger than 8 square pixels.

That area threshold is agent guidance, not a comfortable target size for a person. WCAG 2.2 also has a Level AA target-size criterion of 24 by 24 CSS pixels. WebMCP, mentioned in the web.dev guidance, is a proposed standard in active development with a Chrome origin trial; it is not a requirement for making the controls you already have readable.

<!-- Bad: a clickable div without native button semantics -->
<div class="btn" onclick="book()">Book</div>

<!-- Good: a named button and an associated form label -->
<button type="button" onclick="book()">Book a site visit</button>
<label for="postcode">Postcode</label>
<input id="postcode" name="postcode" autocomplete="postal-code">

WCAG 2.2 success criterion 4.1.2, Name, Role, Value, is an existing Level A standard requiring interface component names and roles to be programmatically determinable. User-settable states, properties and values must also be programmatically settable. Our reading is that this requirement supports the same kind of information the accessibility tree exposes to agents, while serving people who use assistive technology.

The full method for designing websites for AI agents as users extends this check across complete tasks. The first priority is still a control that a person can identify and an agent can inspect.

Check this on your site

In Chrome DevTools, select your main call to action in Elements, then open the Accessibility pane. Check that its role is button or link and that its accessible name describes the action, such as “Book a site visit.”

Myth 5: AI-generated design makes UX skills obsolete

What people say

A design tool can generate a convincing screen from a prompt, so a company no longer needs anyone to make UX decisions. The screen looks finished, and that appearance is easy to mistake for evidence that the task works.

What the source shows

In its 24 October 2025 evaluation of AI prototyping tools, NN/g reported that the tools could follow instructions toward a general goal but lacked the sophistication to weigh design tradeoffs and produce thoughtful, high-quality designs without extensive human guidance. Its 9 May 2025 status update also said that “most design-specific AI cannot replicate human designers’ output quality.”

For a site owner, the design tradeoff is concrete. A generated quote form may look plausible while asking for information your staff does not need, omitting a service area constraint, or leaving the visitor unsure what happens after submission. Generating a screen does not answer those business questions.

Someone still needs to choose the task, establish the facts and test whether the interface helps a person complete it. That work can be done with generated prototypes, hand-built screens or a mixture of both.

Check this on your site

Give a person who has not worked on your site a generated or redesigned page and ask them to request the service they need. Observe whether they can identify the next step and explain what will happen after they submit; write down any point where they stop or guess.

Myth 6: Users trust AI answers, so sites need no trust signals

What people say

Once an assistant has given the answer, a company page only needs to repeat it and provide a button. Visitors will accept the assistant’s price, credentials and availability as settled facts.

What the source shows

In NN/g’s September 2023 diary study, 18 people recorded 425 conversations with ChatGPT, Bard and Bing Chat over two weeks. Only 22.43% of the conversations were followed by verification, and participants checked more often when the stakes were high.

That percentage is a useful warning against assuming every answer gets checked. It is also a reason to make verification possible when the decision warrants it. NN/g studies of nine or 18 people show that behaviors exist, not how common they are across all users.

The February 2026 NN/g study found that several participants distrusted AI-provided pricing and used search to validate details. In a Gartner survey of 645 B2B buyers published on 20 May 2026, 51% said they were more likely to encounter misleading information from generative AI, while 49% said they were more likely to encounter it from a sales representative. Those findings support giving visitors direct access to the facts your company controls.

Put your price or pricing basis, service area, relevant credentials and availability on the page as text. If a quote depends on inspection, say what affects it and what happens next. A visitor can then compare the answer they saw with a statement your company actually maintains.

Check this on your site

Open your main service page and search its text for your price, starting price or explanation of how a quote is calculated. Check whether the visitor can find it without reading text embedded in an image or triggering a hover.

Myth 7: If the AI summary is right, conversion follows

What people say

An AI-referred visitor has already been persuaded. If the assistant gives an accurate answer, the visit should naturally become a booking or inquiry.

What the source shows

Adobe’s March 2025 analysis of more than one trillion visits to US retail sites found that AI-referred visitors were 43% less likely to convert than other visitors in July 2024 and 9% less likely in its March 2025 report. They also had a 23% lower bounce rate.

Later Adobe reports said AI-referred retail visits converted better than non-AI traffic in 2026. Both periods matter: engagement and conversion can move differently, and a retail aggregate does not establish the result for your company. A visitor still needs to understand the offer and complete the form or booking flow.

In GA4, open Reports > Acquisition > Traffic acquisition and use the Session default channel group dimension. Google defines AI Assistant as “the channel by which users arrive at your site from sources like ChatGPT, Gemini, Deepseek, Copilot, or Grok.” Google does not publish the full referrer list.

AI search clicks are counted elsewhere

GA4’s AI Assistant channel does not include clicks from AI Overviews or AI Mode. Those visits count in Organic Search, so the AI Assistant row is not a count of every visit influenced by an AI answer.

Compare the AI Assistant key-event rate with Organic Search for the same period and the same key event. Inspect the session counts alongside the rates: a small number of visits can make a percentage swing sharply. Use the comparison to find pages and tasks to investigate, not to assign every difference to the referral source.

Check this on your site

In the Traffic acquisition report, find AI Assistant and Organic Search under Session default channel group. For the same date range, check the sessions and key-event rate for your primary inquiry or booking event in both rows.

Agent and human readiness check

Run these checks on your most important service page and its booking or inquiry flow. For a company that wants its site legible to assistants and browser agents, Iron Goo offers AI optimization for AI search and agents.

  1. In Chrome DevTools, select the main call to action under Elements and inspect its Accessibility pane. Confirm it has a button or link role and a name that states the action.
  2. Select each form input and compare its accessible name with the visible label. Where an HTML label uses for, confirm that value matches the input’s id.
  3. At 390 pixels wide, inspect the primary button with cookie notices and chat widgets present. Confirm no transparent or visible overlay covers the area you need to click.
  4. Open several pages of the same type and compare the primary button’s position. Confirm it remains in a consistent place as the pages load.
  5. Inspect the HTML for price or pricing basis, service area and booking availability. Confirm the information is present as text and does not depend only on an image or hover.
  6. In Reports > Acquisition > Traffic acquisition, select Session default channel group. Compare the AI Assistant key-event rate and session count with Organic Search for the same period and event.
Free ebook

Put your next steps in order

Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.

By signing up and confirming your address, you consent to receive the free ebook plus one email every two weeks with 3 or 4 new Iron Goo videos, guides and articles, and nothing else; unsubscribe any time with one click.

Related in UX

Ready to move?

Tell us where your business is today. A written assessment lands in your inbox within two business days.

Talk to us