3 GEO strategies that actually worked (and 2 that wasted our time)

We have been optimizing mintec.co for AI search for months. Not as a theoretical exercise — as a real operation, with real traffic, real rankings, and Search Console's AI report as the main dashboard.

We optimized over 430 articles. Some strategies worked well. Others were a waste of time. Here is what moved the needle and what did not.

What we tested

Three main strategies, executed in phases between March and September 2026:

Strategy 1: Restructure existing pages for LLM extraction — headers as answers, atomic paragraphs, the opening paragraph as a complete answer.

Strategy 2: Add specific data and named entities — replacing vague claims like "studies show" with concrete references including source, date, and number.

Strategy 3: Implement FAQPage schema on high-traffic pages — real questions from keyword research, not invented ones.

We also tried two things that did not work: keyword stuffing disguised as "semantic optimization" and adding more content just for volume.

Strategy 1: Restructure for extraction

The biggest change was the simplest to describe and the hardest to execute: rewrite the structure of every article so an LLM can extract a usable answer from any section.

Before, our articles followed the classic SEO pattern: broad introduction, development in long paragraphs with naturally distributed keywords, conclusion. It worked for traditional ranking. It did not work for AI citations.

What we changed:

• Every H2 is now a complete statement or question, not a topic label.
• Paragraphs dropped from an average of 4.2 sentences to 1.8.
• The first paragraph of each article contains the complete answer in 2-3 sentences.
• We added comparison blocks with tables where we previously had descriptive paragraphs.

Result: Restructured pages gained an average of 340% more impressions in Search Console's AI report within 4 weeks. The jump was uneven — pages ranking 1-5 in traditional search gained more than those in positions 6-10.

Takeaway: Structure matters more than content. You can have the best content in the world wrapped in long paragraphs and an LLM will ignore it because it cannot extract a clean answer.

Strategy 2: Specific data beats vague claims

This was the change that surprised us most. We did not expect it to have that much impact.

We replaced generic claims with specific data in 120 selected articles. A real example from mintec.co:

Before:

"Various studies show that structured content performs better in AI search."

After:

"A study by Princeton, Georgia Tech, and IIT Delhi (KDD 2024) found that GEO techniques can increase AI response visibility by up to 40%. On mintec.co, pages with FAQPage schema and specific data are cited 3.2x more by AI Overviews than those with only traditional SEO structure."

Result: Articles with specific data and named entities had 2.1x more brand mentions in manual tests with ChatGPT and Perplexity. This does not appear in Search Console — we measured it by running 50 fixed prompts weekly and recording what each model cited.

Takeaway: LLMs have a bias toward content with verifiable, specific information. They cannot verify the data, but concrete language correlates with authority in their training data. "A 2026 study from Stanford's HAI Institute" sounds more authoritative than "experts say" — and the models treat it that way.

Strategy 3: FAQPage schema as a multiplier

We implemented FAQPage schema on 89 high-traffic pages between April and June 2026. The questions came from real keyword research, not generic ones.

What we did:

• Identified the 5 most searched questions for each article using Google Search Console and autocomplete.
• Wrote self-contained 2-3 sentence answers for each.
• Added FAQPage JSON-LD to each article's frontmatter.
• Verified that answers were extractable separately from the rest of the article.

Result: Pages with FAQ schema had 3.2x more AI Overview impressions than pages without schema in the same period. But there is an important nuance: schema only works if the answers are genuinely useful. We tested 20 pages with generic FAQ answers — no measurable difference.

Takeaway: Schema is a multiplier, not a generator. If your content is weak, schema will not save it. If your content is solid, schema gives the LLM a shortcut to find you.

What did not work

Semantic keyword stuffing. We added synonyms and keyword variations systematically across 40 articles. No measurable difference in AI citations. LLMs do not work like Google crawlers — they do not look for keyword density, they look for useful answers.

More content = better performance. We expanded 30 articles from 1,200 to 2,500 words each. Longer articles did not get cited more. In fact, the longer ones performed worse because they diluted the signal — the answer an LLM was looking for was buried in more text.

The framework we use now

After months of testing, our GEO improvement process comes down to four steps:

1. Audit. We review Search Console's AI report weekly. We identify pages with high impressions that are not being cited — that indicates content Google indexes but LLMs do not choose.

2. Restructure. Every paragraph must be self-contained. If an LLM extracts only that paragraph, does it give a complete answer? If not, rewrite.

3. Anchor. Add at least one specific data point with source per section. It does not matter if it is proprietary or external — what matters is that it is verifiable and named.

4. Measure. Run fixed prompts weekly on ChatGPT, Perplexity, and Gemini. Record what gets cited, what does not, and what changed since last week.

The numbers that tell the full story

After 6 months of optimizing for GEO on mintec.co:

• 60 LLM traffic sessions per week — ChatGPT dominates with 52, Claude grows slowly with 2, Gemini contributes 5-6.
• 3.2x more AI Overview impressions on pages with FAQ schema vs. without.
• 340% more AI impressions on restructured pages vs. their previous version.
• 2.1x more brand mentions in manual prompt tests with specific data vs. vague claims.

None of these metrics appear complete in a single dashboard. Search Console gives you AI Overview impressions but does not tell you what questions generated them. GA4 gives you LLM referral traffic but does not distinguish between an AI Overview click and a direct ChatGPT one. Manual prompt tests give you the most complete picture but require manual work.

The combination of all three is what works. Each one alone tells you half the story.

The question you should be asking

It is not "how do I improve my GEO?" — it is "what am I measuring and what am I missing?"

If you only look at Search Console, you are missing 40% of the picture (external LLM traffic). If you only look at GA4, you are missing another 40% (AI Overview impressions that do not generate clicks). If you are not running manual prompt tests, you are missing the most important layer: what AI says about your brand when someone asks directly.

GEO is not SEO with a different name. It is a new discipline with new tools, new metrics, and tactics we are still learning. What worked for us might not work for you — but the process of testing, measuring, and adjusting is universal.