What Happened
HubSpot published a piece on 21 September 2026 arguing that answer engines show their reasoning for psychological reasons, not purely transparency ones. It notes that in 2025 most of the major answer engines made an almost identical update: Claude, ChatGPT, Gemini and many others started to show what they were thinking.
HubSpot frames the change around load time. A visible reasoning trace fills the gap between the moment you press send and the moment an answer appears, and that gap is where people abandon interfaces.
The underlying principle is well established in service design. People judge a wait more favourably when they can see work happening, which is why progress indicators, skeleton screens and staged loading states have been standard user experience practice for years. What changed in 2025 is that the output of the work became the waiting state itself. The model’s intermediate steps serve two purposes at once.
One point HubSpot’s piece does not cover: this creates an awkward split for UK site owners. Google’s Core Web Vitals still measure real timings, including Largest Contentful Paint and Interaction to Next Paint, and they do not award credit for a convincing loading animation. Perceived speed wins users. Measured speed wins the technical audit. You need both.
Applying perceived performance to your own site
Start by finding every point on your site where a user waits without feedback. Form submissions, filtered product listings, booking calendars and quote calculators are the usual offenders. If any of them show nothing for more than a second, you have an abandonment problem that no amount of copywriting fixes.
Then decide what the waiting state should say. Generic spinners tell the user nothing. A message naming the step, for example checking availability or calculating your quote, borrows exactly the mechanism the answer engines adopted. This costs a developer a few hours, not a redesign, and it is one of the few user experience changes where the improvement is visible in session recordings within a fortnight.
If you are adding an AI tool to your own site, whether a support assistant or a recommendation feature, stream the response rather than holding it back until complete. Users tolerate a slow answer that is visibly arriving far better than a fast one that appears after a blank pause. Teams building this into client sites can see how we handle it in our AI services, where the interaction design matters as much as the model behind it.
Perceived speed is not a substitute for real speed
Dress up a four-second wait and you still have a four-second wait, one that Core Web Vitals will report accurately. Fix the server response and the image weights first, then use the honest remaining delay to communicate rather than to stall.
How we approach AI interfaces on client sites
We treat any customer-facing AI tool as a user experience project with a model attached, not the other way round. Before writing a prompt, we map the wait points, the failure states and what the interface says when the model returns nothing useful. That last one gets skipped most often and causes the most support tickets.
We are also direct about measurement. An AI tool on a landing page should be judged on completed enquiries and task completion rate, not on conversation volume. We set those baselines before launch so the review at 8 to 12 weeks has something real to compare against.
Questions about AI interfaces and perceived speed
Is a reasoning display worth showing on our own chatbot?
Only if there is genuine processing to narrate. A fake thinking animation on an instant response reads as manipulative once users notice the timing never varies.
Will improving perceived speed help our Google rankings?
No. Core Web Vitals measure actual timings, so perceived improvements help conversion and retention while genuine performance work helps the technical picture.
How long does it take to add streaming responses to an existing site?
For a standard implementation, expect a developer to need days rather than weeks, though integration with an older CRM or booking system can extend that considerably.
Thinking through perceived speed on your site
If your analytics show people dropping out at a specific interaction and you are not sure whether the cause is speed, clarity or trust, that is a useful conversation to have with someone outside the build. You are welcome to talk to our team about what your data is actually showing.
Written by the London Marketing Company team.
Source: https://blog.hubspot.com/marketing/why-ai-shows-loadtime
