Google Is Adding More Links to AI Search. Your Website Still Has to Earn the Click

Google has started changing how links show up inside AI Mode and AI Overviews.

That is useful. It is also not a reason to relax.

If anything, it makes the practical job clearer: your website needs to be easy for AI systems to crawl, understand, quote, preview, and send humans back to. Not in a gimmicky "rewrite everything for AI" way. In the boring-but-important way: clear pages, stable signals, honest metadata, visible evidence, and content that can be understood without guessing.

This is the bit I think matters most for normal businesses: AI search may be adding more routes back to websites, but those routes still need somewhere worth landing.

What changed

On 6 May 2026, Google said it is rolling out new AI Mode and AI Overview features that show more ways to explore the web from AI answers.

The interesting parts for site owners are:

• AI responses can show follow-up links to deeper articles and original sources.
• Some links can appear inline next to the specific point they support.
• AI answers can include previews of websites before someone clicks.
• Google is adding more context around firsthand perspectives from forums, communities, and social sources.
• Subscription publishers may get clearer subscription-labeled links when users have connected access.

In plain English: Google is trying to make citations and click paths inside AI answers feel less like a footnote and more like part of the answer.

That is a good direction for the open web. But it does not mean every useful page suddenly benefits.

Two things can be true at once

The first truth: better links in AI search are good for publishers, SaaS companies, local businesses, and anyone who has useful original content.

The second truth: answer-first search still changes behaviour.

A recent working paper looking at Google AI Overviews and Wikipedia traffic estimated that AI Overview exposure reduced daily traffic to English Wikipedia articles by about 15%. That does not mean every site should expect the same number. Wikipedia is not your B2B SaaS site, ecommerce store, agency site, or local service business. But it is a useful warning that AI answers can satisfy some intent before a click happens.

Another recent study comparing Google Search, AI Overviews, and Gemini found that AI Overviews appeared for 51.5% of representative real-user queries in its dataset, and that source sets differed substantially between traditional and generative search. It also found that sites blocking Google's AI crawler were less likely to be retrieved by AI Overviews.

So the practical takeaway is not "AI search is dead for websites" or "just allow every bot and hope".

The takeaway is more grounded:

Your content has to be both useful enough to be cited and technically easy enough to be used.

This is where most websites are still early

Cloudflare's recent Agent Readiness work scanned 200,000 high-traffic domains and found that:

• 78% had a robots.txt file, but most were written for traditional crawlers rather than agents.
• 4% declared AI usage preferences through Content Signals.
• 3.9% supported Markdown content negotiation.
• MCP Server Cards and API Catalogs were barely visible in the dataset.

OpenRobotsTXT is also now tracking AI content-usage adoption. At the time of writing on 14 May 2026, it showed 1,763 root domains and 3,770 hostnames using the newer AI content-usage proposal signals.

Those numbers will move. But the current shape is pretty obvious: most sites have the old crawl basics, far fewer have made their content and policies genuinely agent-readable.

That gap is the opportunity.

Not because every small business needs an MCP server tomorrow. Most do not.

But because a large number of websites still fail the basics that make a page trustworthy and reusable:

• The useful answer is buried below generic hero copy.
• The page depends on client-side rendering before key text appears.
• The title and description do not match the real page.
• The canonical, sitemap, robots, and llms.txt signals disagree.
• The WAF blocks normal crawler behaviour while the public policy says crawling is allowed.
• The page has no date, no author context, no pricing clarity, no product state, or no evidence.

AI search makes those weaknesses more visible because it works at the level of chunks, claims, and citations.

A practical checklist for this new AI search shape

Here is the version I would actually give a team. It avoids the theatre and focuses on what is most likely to help.

1. Make the important answer visible early

If a page exists to answer a question, answer it clearly.

Put a short, direct explanation near the top. Then expand below it. This helps humans, search engines, and AI systems.

Bad pattern:

• 900 words of brand positioning before the answer.
• Key details hidden in accordions that only hydrate after JavaScript.
• Important facts only shown in images or decorative cards.

Better pattern:

• One clear H1.
• A short intro that states the practical answer.
• Descriptive H2s.
• Specific examples, caveats, and verification steps.

This is not "write for robots". It is just better publishing.

2. Keep crawl signals aligned

Most visibility problems are not dramatic. They are boring contradictions.

Check:

• robots.txt says the page can be crawled.
• The page is included in the sitemap if it is important.
• The canonical URL points to itself or the real canonical page.
• The meta robots tag is not accidentally noindex.
• The page returns 200, not a soft error.
• AI crawler rules do not conflict with llms.txt or Content Signals.

One contradiction may not kill a page. Several contradictions make it harder for systems to trust what you meant.

3. Add machine-readable policy where it genuinely helps

For many sites, the sensible baseline is:

• robots.txt for crawl access and sitemap discovery.
• Content Signals if you want to state AI usage preferences.
• llms.txt if you have a clear AI-facing summary of important public resources.
• Markdown content negotiation for documentation, help, product explainers, or guides where clean text helps.

Do not add files just to look modern. Add them if you can keep them accurate.

A stale llms.txt is worse than no llms.txt because it gives agents a confident wrong map.

4. Make claims easy to verify

AI answers often quote or summarize small parts of a page. That means weak claims age badly.

For pages that matter, add:

• Dates where freshness matters.
• Author or organisation context.
• Real examples.
• Screenshots or diagrams when they clarify the point.
• Clear product limits and pricing.
• Source links when you reference external data.

The goal is not to add fake authority. The goal is to make real authority easier to see.

5. Watch server logs, not just analytics

If AI systems reduce some click-through, analytics alone can understate what is happening.

You want to know:

• Which bots are crawling.
• Which paths they request.
• Whether they hit 200, 403, 404, or redirects.
• Whether your edge rules block crawlers you intended to allow.
• Whether AI bot access matches your written policy.

This is where a lot of teams find the uncomfortable truth: the policy says "allowed", but the edge says "blocked".

6. Measure content usefulness after the click

If Google is adding richer previews and inline links, the click becomes more qualified.

That means the landing page has to do its job quickly.

For commercial pages, check:

• Does the page answer the query that probably generated the citation?
• Is the next step obvious?
• Is pricing or contact context clear?
• Does the page load quickly on mobile?
• Does the user land on a stable section, not a jumping layout?

AI search does not remove normal UX work. It makes weak UX more expensive because fewer clicks may make it through in the first place.

What not to do

Do not panic-rebuild your site because AI search changed again.

Do not publish a pile of thin "AI visibility" pages that say nothing new.

Do not fake reviews, ratings, authorship, or product data because a schema tool said a field was "recommended".

Do not assume robots.txt is access control.

Do not assume an AI crawler rule is working until you have tested the actual response from outside your own machine.

The best response is still the grounded one: make useful pages easier to discover, parse, trust, and revisit.

What to do next in Scavo

If you want to sense-check this without turning it into a month-long project:

1. Run a fresh scan on your homepage, pricing page, and one important article or guide.
2. Check the AI visibility section first: crawler access, robots.txt, llms.txt, Content Signals, markdown support, and policy conflicts.
3. Check the SEO basics next: title, description, canonical, sitemap inclusion, structured data, and indexability.
4. Open the linked help guide for anything failing and assign the smallest safe fix.
5. Re-scan after deployment and confirm the live page says what you think it says.

The goal is not to chase every new standard. The goal is to stop leaving unclear signals in places where humans, search engines, and AI systems all now have to make decisions about your content.

Sources

• Google: 5 new ways to explore the web with generative AI in Search
• Cloudflare: Introducing the Agent Readiness score
• Khosravi and Yoganarasimhan: Impact of AI Search Summaries on Website Traffic
• Grossman et al.: How Generative AI Disrupts Search
• OpenRobotsTXT AI content-usage adoption stats
• Checkly: The Current State of Content Negotiation for AI Agents