How Google's AI Mode Changes Your Site: Task-Based Search, AI Overviews and Falling CTR

Summary: Google search is changing in three directions at once: queries are moving from a single question to a multi-step task, AI Overviews give the answer on the results page, and users increasingly move over to AI tools. In analytics the result looks the same: impressions grow much faster than clicks, and CTR falls. In this article we show the change with our own Search Console data and give seven steps to prepare your site.

Three changes happening in search at the same time

What has happened over the past year can't be gathered under one heading, because it is not a single update but three trends overlapping. Each of them touches your site's place on the results page in a different way.

1. The query becomes a task

Google search is no longer just a tool for giving information. It is developing into a system that helps the user complete a whole task from start to finish. This is called task-based search.

An example. A user doesn't simply think "buy office furniture in Baku". The sequence goes like this: choose, compare, find out the delivery time, check the payment terms, place the order. If Google can close several steps of this path on the same site, it prefers that site. The other side of it is contextual continuity: earlier queries in the session are taken into account, so your internal link structure has to answer that context. A lone page that links nowhere is a weak signal here.

2. The answer is given on the results page

An AI Overview is a summary prepared by artificial intelligence at the top of the results. When users get the answer there, they don't go on to the site. The result is measurable: your site is shown and even cited, but the rate at which impressions turn into clicks falls. This is not a penalty; it is a change in the results page.

3. Users move to AI tools

ChatGPT, Perplexity, Gemini and Google's own AI Mode are already the starting point for many queries. So your site has to be readable not only for Google's bot but also for the models that gather answers. If the content is not in a question-and-answer format and doesn't state facts clearly, the chance of being chosen as a cited source is low.

How we saw this in our own data

So that it doesn't sound theoretical, we show our own Search Console figures openly:

These figures are often misread. Clicks did not drop; they grew 15 percent. But impressions grew 47 percent, three times faster than clicks, and CTR fell from 2.15 percent to 1.68 percent. The site is shown more, and a smaller share of those impressions turns into clicks. If the same scissors open in your report, the problem is not that your site "got worse" but that the results page changed.

The second figure is about devices: mobile CTR is 1.39 percent, desktop CTR 2.18 percent. Mobile brings more impressions, but clicks come harder there. Mobile experience here is not a matter of design taste but directly a matter of clicks.

The third figure answers the quantity approach. The 356 posts on our blog brought 81 clicks in total over the last 90 days, and 46 of them went to a single post. The number of posts doesn't deliver results. One deep post that answers the question to the end does more than dozens of shallow ones.

How to run the same check on your own report

1. In the Performance report, choose "Compare" in the date filter and set the last 90 days against the previous 90 days.
2. Turn on all four metrics together: clicks, impressions, CTR and average position. Looking at clicks alone is misleading.
3. Compare the growth rate of impressions with the growth rate of clicks. If impressions grow much faster, the scissors have opened.
4. In the Queries section, separate the searches made with your brand name. The rest is your real non-brand demand.
5. In the Devices section, read mobile and desktop CTR separately. If the gap is large, the work lies in the mobile experience.
6. In the Pages section, list the pages whose impressions grew while their clicks didn't. Start your fixes with them.

The old model and the new one

• The old model: the user asks, Google gives a list of results, the user goes to a site and decides on their own.
• The new model: the user asks, the system analyses the task, gives part of the answer itself and sends the user only to sites with a clear structure that it trusts.

The difference looks small, but its effect is big: if your structure, speed and content architecture are not in a format a machine can read, you are outside this chain.

Which sites lose less in this environment

• Sites that are strong in brand search. When the brand name is typed directly, the AI summary doesn't stand in between.
• Sites with deep, structured content. Cited sources are chosen from exactly such pages.
• Sites with strong E-E-A-T signals. When experience, expertise, authority and trust are visible on the page, the chance of being chosen as a source grows.
• Sites with a healthy technical base. Speed, semantic HTML, a canonical URL structure and structured data.

Is your site ready: six questions

• Does it open quickly on a mobile connection? The LCP target in Core Web Vitals is under 2.5 seconds.
• Does everything work on mobile, including forms and the payment flow?
• Is there JSON-LD structured data? FAQ, HowTo and BreadcrumbList are the most useful.
• Is the heading hierarchy logical, or are headings used just as a font size?
• Is SSL active, and are all addresses gathered into a single canonical version?
• Are your service pages linked to each other and to the posts that give depth?

If you answer "no" to any of them, you know what goes at the top of your fix list.

What to do: seven steps

1. Turn intent into a task map

For every service and product page, ask: after reading this, what is the user's next step? Show that step clearly on the page. A comparison table, a price block, a clear CTA and a short FAQ do this job.

2. Build a pillar and cluster structure

One main page with narrower topics around it, all linked to one another. This structure builds topical authority and shows the models that gather answers which field your site goes deep in.

3. Add structured data

Use FAQ, HowTo and BreadcrumbList markup to tell the machine the steps of the task explicitly. This enriches how you look in the results and makes it easier for the content to be understood correctly.

4. Strengthen brand search

It is the least talked-about item, yet the one that works best. When a user searches for a brand name directly, the AI summary doesn't stand in between, and the click comes to you. We see this from the other side in our own data: most of the 189 clicks in our query report go not to our own brand but to the brand names of the sites we have built. With client brands removed, our own queries have 49 clicks left. Brand search brings clicks; it is just that the brand is our clients', not ours. Your job is to build that brand in your own name.

5. Write in answer format

The texts generative systems cite are built the same way: the question in the heading, the answer in the first sentence, the explanation after. Long introductions and texts that reach the topic only after five paragraphs get lost in this system. This approach has become a field of its own under the names AEO (answer engine optimization) and GEO (generative engine optimization), but the core is the same: don't hide the answer.

6. Show who wrote it

This step is not in the code; it is in editorial discipline. Every post should carry the author's name and expertise, plus the publication and last-update dates. If you use a number, say where it comes from. On the company page, show the team, the address and a contact channel. E-E-A-T is nothing more than this: can the page say who wrote it, when, and on what basis. A text without a name, a date and a source is anonymous to readers and models alike.

7. Don't leave speed and mobile experience for last

Speed and mobile friendliness are not decorations added at the end of a project; they are architectural decisions. Patching them later is expensive and never gives the full result. That is why these requirements should be written down from the moment you order a website.

What to look for when ordering a website in Baku

Most businesses look only at the design when they order a website. The technical structure, speed and internal linking fall into the background, and the result is a site that looks good but doesn't show up in search. Ask about these in the conversation: who writes the code, what the speed target is, whether structured data is included and how many months support lasts.

SaytPro is a web studio based in Baku, founded in 2023, and a real team of developers. More than 120 completed projects, 3 years of experience and more than 60 active clients. Stack: Next.js, TypeScript, Tailwind, Django, PostgreSQL, Hetzner, Cloudflare.

Prices are public: a landing page starts at 800 manat, a corporate website at 1,300 manat and an online store at 1,800 manat. SEO is 700 manat a month, SMM and advertising 600 manat a month, a chatbot from 900 manat, UI/UX from 800 manat. The price includes the domain, 1 year of VPS hosting and SSL. After handover there are 2 months of formal warranty and technical support, and from the second year hosting is 120 manat a year. Payment is 50 percent at the start and 50 percent on handover. The first consultation is free.

Frequently asked questions

If impressions grew and CTR fell, has my site been penalised?

No. A penalty usually brings impressions themselves down. If impressions grow while CTR falls, then for some of your queries the answer is already given at the top of the page. The fix lies not in rankings but in titles, descriptions and the format of the content.

Is it possible to get clicks at all on a query with an AI summary?

Yes, depending on the query. Queries with brand, price, comparison or clear purchase intent still bring clicks. On general definition queries the share of clicks shrinks, so the content plan should shift towards queries with commercial intent.

Does task-based search need a separate technology?

No. What it needs is the structure of your existing site: a clear heading hierarchy, structured data, internal links and a next step shown on every page. These are ordinary technical SEO tasks; they simply matter more now.

Does writing lots of blog posts help?

Our own experience shows the opposite. 356 posts brought 81 clicks in 90 days, and more than half of them went to a single post. Deepening existing posts gives more results than adding to their number.

Is it better to fix the existing site or rebuild it?

The answer depends on the technical base. If speed, mobile experience and the heading structure can be fixed, working on the existing site is cheaper. On template sites whose code can't be touched, every fix makes the next one harder, and there a rebuild pays off.

Next step

Open your Search Console report and run the six-step check above. If impressions grow faster than clicks and CTR falls, the seven steps in this article are your list. If you don't know where to start, the SEO service page shows how the work is set up, and the price page helps you work out a budget. If you are planning a new site, the corporate website page shows where we start.