AI Overviews and Your Blog: How to Stay Visible (GEO Basics)

Search Console has been showing me the same pattern for months: impressions climb, clicks stay flat. The post still ranks. People still see it in results. They just don't click anymore, because Google's AI Overview answered their question before they reached the blue links.

If you run a blog, you've probably seen some version of this. It's not a glitch, and it's not going away. Google now shows AI Overviews on a large share of informational queries, and a growing slice of your audience never opens Google at all — they ask ChatGPT, Perplexity, or Claude instead.

TL;DR: AI systems answer informational queries directly, so clicks on those queries drop. The response isn't panic — it's generative engine optimization (GEO): structure content so AI systems can cite it, keep your site machine-readable, decide deliberately whether AI crawlers may access it, and shift your writing toward things a summary can't replace. This post walks through each step.

What Changed: The Answer Arrives Before the Click

Two shifts happened at once, and they compound each other.

First, Google answers queries on the results page. AI Overviews — the generated summary above the organic results — now appear on a meaningful share of informational searches. Pew Research tracked real user behavior in 2025 and found that when an AI summary appeared, users clicked a traditional result on roughly 8% of those searches, against about 15% when no summary appeared. That's close to half the clicks gone on affected queries.

Second, chat assistants became search engines. ChatGPT, Perplexity, and Claude all browse the web and answer questions with synthesized text plus a handful of source links. For a chunk of users — especially developers — "search" now means typing a question into a chat box.

The damage isn't evenly spread. Queries like "what is a headless CMS" or "how does JSON-LD work" get fully answered in the Overview. Queries with intent behind them — comparisons, purchase research, "best X for Y" — still produce clicks, because people want to verify before acting.

Here's the part that matters: AI answers cite sources. Google's Overviews link out. Perplexity footnotes nearly every sentence. ChatGPT shows sources when it browses. Being one of those cited links is the new position one — and it's something you can influence.

What Is Generative Engine Optimization (GEO)?

Generative engine optimization is the practice of making your content easy for AI systems to retrieve, understand, quote, and cite. Where classic SEO targets a ranked list of links, GEO targets the synthesized answer — the goal is to be the source the machine quotes, with your name attached.

The honest version: GEO is about 80% the same work as good SEO. Clear structure, direct answers, real expertise, clean technical foundations — Google has rewarded all of that for years, and AI systems inherited the preference because they were trained and grounded on the same web.

The remaining 20% is new. It covers:

1. Citation-worthiness — writing passages an AI can lift cleanly and attribute.
2. Machine readability — clean HTML, structured data, and emerging conventions like llms.txt.
3. Crawler policy — deciding which AI bots may read your site, and knowing the tradeoff.
4. Different measurement — citations and referrals instead of rank positions.

None of this replaces SEO. A page that can't rank usually can't get cited either, because AI search systems lean on conventional search indexes to find candidate sources in the first place.

What AI Systems Actually Cite

I've spent a lot of evenings asking Perplexity and ChatGPT the questions our own posts target, then checking which sources they pull. Patterns repeat. Four things show up again and again in cited pages.

Direct answers near the top of each section

AI retrieval works on passages, not whole pages. When a system needs to answer "should I block GPTBot," it pulls the paragraph that answers it — not your 300-word wind-up about the history of web crawlers.

So front-load. After every H2, give the answer in the first one or two sentences, then expand. If a section's first paragraph could stand alone as a quote in an AI answer, you've done it right.

Descriptive headings and predictable structure

Headings are retrieval signals. A heading like "Thoughts on the Bot Situation" tells a machine nothing. "Should You Block AI Crawlers in robots.txt?" tells it exactly what the section answers, and matches the way people phrase questions to chat assistants.

Lists and tables help for the same reason: they're unambiguous. A comparison table doesn't need interpretation. A numbered list of steps maps directly onto a "how do I" answer. Pages that mix prose with structured elements get cited more than walls of text, in both published studies and my own spot checks.

Original data and first-hand experience

AI systems synthesize. What they can't do is generate a number that doesn't exist anywhere. If your post contains an original benchmark, a survey you ran, real traffic figures, or the results of testing eight tools yourself, an AI answer that wants that information has exactly one source to cite: you.

Derivative content faces the opposite problem. If your post says the same thing as forty other posts, the machine can take the consensus from anywhere and cite anyone — usually the biggest domain. Being the origin of a fact is the strongest citation magnet there is.

Consistent entity information

AI systems track entities — your product name, your company, your authors — across the web. If your site says one thing and your GitHub, directory listings, and social profiles say another, confidence in citing you drops.

Keep it boring and consistent: same product name everywhere, same one-line description, same author names with real bios. Structured data helps here too, because JSON-LD states entity facts in a format machines parse without guessing.

Make Your Pages Machine-Readable

A page gets cited only if the crawler can read it cleanly. Server-rendered HTML, JSON-LD structured data, an accurate XML sitemap, and an llms.txt file together make your site legible to AI systems with no JavaScript execution and no guesswork.

This is the technical half of GEO, and it's mostly solved at the platform level.

Clean, server-rendered HTML. Most AI crawlers don't execute JavaScript. GPTBot doesn't. ClaudeBot doesn't. If your content only exists after a client-side framework hydrates, those bots see an empty shell. Server-rendered pages — the default for a traditional CMS like UnfoldCMS, which is a self-hosted Laravel application shipping fully rendered HTML and automatic Article JSON-LD — have a real advantage here over client-heavy stacks. If you're on a JS-first setup, pre-render your content pages.

JSON-LD structured data. Article schema tells machines the title, author, dates, and publisher without scraping the layout. It's also where entity consistency lives. If your platform doesn't generate Article schema automatically, that's worth fixing this quarter — our CMS SEO checklist covers structured data and the other features worth demanding from any platform.

llms.txt. An emerging convention: a markdown file at your domain root that tells language models what your site is and links to your most useful pages in a clean, parseable format. Adoption by AI vendors is still early and uneven, but it costs almost nothing to maintain and the spec is gaining mindshare. We've written a full explainer on what llms.txt is and how to set it up.

XML sitemap. Nothing new, still required. AI search systems discover URLs the same way Google does.

Should You Block AI Bots in robots.txt?

There's no free answer here. Blocking AI crawlers keeps your content out of training data and AI answers alike. Allowing them gives you citation visibility but means your writing feeds systems that sometimes answer the question without sending you the click. Decide based on what your content is for — and know the bots by name first.

Two details trip people up.

You can't opt out of AI Overviews via robots.txt without leaving Google Search. AI Overviews use the regular Googlebot crawl. Blocking Google-Extended only opts you out of Gemini training. The only way out of Overviews is noindex — which removes you from search entirely. That decision is effectively made for you.

Training bots and search bots are different animals. Blocking GPTBot stops your content joining future training sets but OAI-SearchBot still powers live ChatGPT citations. You can block one and allow the other.

My take, for a blog whose job is to be found: allow the search-focused bots (OAI-SearchBot, PerplexityBot, ClaudeBot) and make the training-bot call based on principle, because the traffic effect of blocking GPTBot alone is small. If your content is your product — paid courses, proprietary research you sell — blocking everything is rational. Both positions are defensible. Drifting along with whatever your robots.txt happened to say in 2023 is not.

Whatever you choose, you need easy control over the file. UnfoldCMS includes a robots.txt editor in the admin, so adding or removing AI bot rules is a settings change rather than a server deploy. On other platforms, check where the file lives before you need to change it quickly.

Write Content That Survives the Summary

If an AI Overview can fully replace your post, the post was always vulnerable — Wikipedia and a thousand rewrites of the same definition were already competing for it. The durable move is shifting your mix toward content a summary can't replace:

• Opinion with a name attached. "Here's why I think llms.txt will matter" can be quoted, but readers click through for the full argument. Machines synthesize consensus; they can't hold a position.
• First-hand experience. Migration war stories, benchmarks you ran, screenshots of real dashboards. An AI can repeat your findings — with a citation — but it can't have been there.
• Comparisons with judgment calls. A spec table gets absorbed into the answer. "Tool A is better until you hit 50k entries, then the editor chokes" is a verdict people want to verify at the source. Our Contentful alternatives comparison earns its clicks from exactly those judgment calls, not from the feature matrix.
• Tools and products. An AI can describe your product. It can't be your product. Content attached to something usable — a template, a calculator, an open-source repo — keeps its reason to be visited.

Definitional content still has a job: it earns citations, builds topical authority, and feeds the entity signals everything else rests on. Just stop expecting it to deliver the clicks it did in 2022, and stop making it the whole strategy.

How to Measure What AI Search Is Doing to You

You can't manage what you don't measure, and AI search effects hide in plain sight inside tools you already have.

1. Impressions vs. clicks in Search Console. Filter to informational queries. Impressions holding or rising while clicks sag is the AI Overview signature — you're being seen, and increasingly being read by a machine instead of a person. Track CTR per query cluster quarter over quarter, not week to week.
2. AI referral traffic in your analytics. Look for referrers from chatgpt.com and perplexity.ai. The volumes are small for most sites today, but they're growing, and these visitors arrive pre-qualified — someone clicked through from an answer that already mentioned you.
3. Manual citation checks. Once a month, ask ChatGPT (with browsing), Perplexity, and Google the ten questions your most important posts target. Note whether you're cited and who is. Crude, but it's the closest thing to rank tracking AI answers have right now.
4. Server logs for AI crawlers. Grep your access logs for GPTBot, ClaudeBot, and PerplexityBot user agents. If they never visit, your robots.txt, firewall, or CDN may be blocking them without you having decided that.

What NOT to Do: Bulk AI Content

The tempting response to AI search is using AI to mass-produce posts. It's also the single most punished tactic in modern search.

Google's spam policies target "scaled content abuse" by name, and the recent core updates hit mass-produced, derivative content harder than anything else — sites running that playbook have seen traffic losses of 50% and worse. The irony is sharp: AI search engines demote AI-generated bulk content hardest, because synthesized rewrites of existing pages add nothing for a system whose whole job is synthesis. You'd be feeding a summarizer more summaries.

Drafting assistance is fine; nobody can tell and nobody penalizes it. Publishing forty unedited posts a week with no original information is the part that ends badly. The bar AI search sets is being a primary source. Volume without information moves you in the opposite direction.

FAQ

Does blocking GPTBot remove my site from ChatGPT?

Not from live search answers. GPTBot gathers training data; OAI-SearchBot powers ChatGPT's browsing and citations. Block GPTBot and allow OAI-SearchBot, and ChatGPT can still cite your pages in real-time answers while your content stays out of future training runs.

Is GEO a replacement for SEO?

No. AI systems find candidate sources through conventional search indexes, so pages that can't rank rarely get cited. Treat GEO as a layer on top of SEO: same foundations, plus citation-friendly writing, machine readability, and a deliberate crawler policy.

Will AI Overviews kill blog traffic entirely?

Unlikely. Click-through on purely definitional queries has fallen and will keep falling, but comparison, opinion, and decision-stage queries still drive clicks because readers verify before they act. Blogs weighted toward experience and judgment are holding up; blogs built on rewritten definitions are not.

Where to Start

If you do nothing else this month: front-load direct answers under every H2 of your ten most important posts, confirm your pages render as full HTML without JavaScript, ship JSON-LD and an llms.txt file, and make a deliberate robots.txt decision instead of inheriting an accidental one. Then watch impressions vs. clicks in Search Console and adjust from evidence, not headlines.

None of this is exotic. It's the same discipline good SEO always demanded, aimed at a second audience that reads HTML instead of pixels. The sites that lose this transition will mostly be the ones that kept publishing interchangeable content and never noticed the rules changed.

Sources: Pew Research Center's 2025 study of search behavior with AI summaries; Google Search documentation on crawler user agents and spam policies; OpenAI, Anthropic, and Perplexity crawler documentation; our own Search Console data for unfoldcms.com. This post is published on a CMS vendor's blog — we build the kind of server-rendered, structured-data-first platform this article recommends, so weigh our take accordingly.