3 GEO experiments you should try this year

Final month, I requested ChatGPT, Perplexity, and Gemini the identical query about three of my purchasers: “Who’s [Brand Name] and what do they do?”

Two out of three acquired it mistaken. Unsuitable companies. Outdated workplace places. One even steered a competitor as a greater different.

Right here’s what makes that greater than a curious mistake.

What in the present day’s AI errors reveal about model visibility

AI-sourced traffic jumped 527% year-over-year from early 2024 to early 2025. 

Whereas that development is actual, it’s rising from a really small base. Most websites nonetheless see AI referrals representing lower than 1% of whole visitors. 

However when half the AI-generated descriptions of your model are inaccurate, that’s not only a future drawback. That’s shaping perceptions proper now.

The problem isn’t whether or not to optimize for AI programs – it’s how to take action successfully. 

It’s determining what truly works versus what’s simply repackaged fundamentals being bought as one thing revolutionary. 

And in contrast to conventional SEO, the place we will forecast visitors and income with affordable confidence, AI search doesn’t work that means. 

You possibly can’t promote certainty right here. You possibly can solely promote managed studying.

Best GEO ways grow to be web optimization fundamentals utilized to a brand new visibility layer. 

Construction, readability, and constant data have at all times mattered. 

What’s modified is that these ideas now influence how AI programs summarize and cite your content material, not simply how customers discover and work together with it.

The one approach to separate reality from fiction is to run small, reversible experiments that produce decision-quality information. 

The price of not figuring out what works is increased than the price of discovering out.

Beneath are three GEO experiments you’ll be able to run to know how AI programs learn, summarize, and reuse your content material. 

These are sensible exams most groups can full in 60–90 days, and each produces clear insights about whether or not these ways truly transfer the needle for what you are promoting.

Consider these as managed studying alternatives, not visitors guarantees. 

Experiment 1: Construct an LLM-ready subject cluster

Entrepreneurs have been constructing topic clusters for years. However GEO modifications the principles.

Generative programs don’t learn content material the best way people do. 

They chunk it, on the lookout for clear entities, clear solutions, constant language, and predictable construction. 

When your content material is organized on this means throughout a complete cluster, it turns into simpler for AI programs to know and cite you as a most well-liked supply.

This primary experiment exams precisely that.

Choose a cluster with enterprise worth

Select a subject the place you have already got robust content material or the place you desperately must develop visibility.

Use inner web site search, Google Search Console queries, and buyer help calls to seek out the natural-language questions your viewers is already asking. 

These are sometimes comparable queries or prompts potential prospects use in LLM platforms.

Tip: In case your help staff hears the identical query thrice in per week, that’s your sign.

Construct (or rebuild) the cluster for machine readability

Right here’s what I’ve seen work throughout exams.

• Construction your pillar web page round natural-language questions
  • Your H2s ought to mirror the best way actual people phrase queries:
    • “What’s [topic]?”
    • “How a lot does [topic] price?”
    • “What’s the best choice for newbies?”
    • “What ought to I keep away from?”
  • AI instruments favor pages that reply questions in the best way customers truly ask them. Not the best way we suppose they need to ask them.
• Lead with a summary-first design
  • Make the primary 100–150 phrases a quick, clear overview.
  • No gradual intros. No storytelling wind-up. No “In in the present day’s fast-paced digital panorama…” fluff.
• Use constant Q&A formatting
  • Break down each web page with predictable formatting:
    • Query
    • Brief reply (1-2 sentences)
    • Supporting element (2-3 paragraphs)
    • Elective desk or record
  • This format is nice for LLMs. It tells them precisely the place to look and what to extract.
• Don’t skip schema and inner hyperlinks
  • Use FAQPage, HowTo, Product, Group, LocalBusiness, no matter’s related to your content material.
  • Use inner hyperlinks to determine the cluster hierarchy so fashions don’t should guess which web page solutions which query. Make the connection between the pillar and supporting pages clear.

Measure the correct issues

Right here’s what to trace over the subsequent 60 days:

• AI Overview appearances on your goal cluster queries (use incognito mode and test manually twice per week, or use instruments like Semrush you probably have entry)
• LLM quotation patterns: Run the identical queries by ChatGPT, Gemini, and Perplexity. Do they reference your web site? How precisely?
• Natural visitors and conversions inside the cluster
• Consistency of descriptions: Are LLMs describing your content material the identical means, or are they confused?

Right here’s the important thing distinction.

In conventional web optimization, we concentrate on visitors and income metrics. 

With GEO experiments, you’re monitoring main indicators, indicators that inform you whether or not AI programs perceive and belief your content material, even earlier than these indicators translate into measurable visitors.

Consider it like this: Quotation accuracy and entity recognition are the brand new “rankings.” 

They point out whether or not you’re positioned to profit as AI search quantity grows.

Evaluate in opposition to a management

That is essential: check this cluster in opposition to one other one you didn’t optimize.

If the LLM-ready cluster reveals extra AI Overview inclusion, extra correct solutions, and steadier natural efficiency, you’ve discovered a lever price scaling.

Instance: 

• I rebuilt a subject cluster for a dental apply round “tooth whitening choices.” 
• Inside 75 days, they appeared in AI Overviews for 9 out of 13 goal queries, up from two. 
• Conventional natural visitors held regular, however the model’s visibility in AI-generated solutions elevated.

Why this works (past simply AI)

Right here’s what makes this experiment notably invaluable: the identical structural enhancements that assist AI programs perceive your content material additionally have a tendency to enhance conventional search efficiency.

Clear headings, direct solutions, and logical content material group assist Google parse your content material extra successfully. 

Customers recognize the readability, too. Shorter time to discovering solutions sometimes correlates with higher engagement metrics.

So even when AI visitors stays a small proportion of your whole visitors, you’re constructing content material that performs higher throughout all channels.

That’s the form of optimization price investing in.

Dig deeper: Chunk, cite, clarify, build: A content framework for AI search

Experiment 2: Run a model entity and sentiment dash

AI is horrible at nuance.

In case your model story isn’t constant throughout platforms, LLMs will generally make one thing up, or worse, they’ll confidently inform customers one thing utterly mistaken about you.

Fashions pull model data from:

• Critiques (Google, Yelp, Trustpilot, area of interest directories).
• Enterprise directories.
• Editorial content material and information mentions.
• Reddit and business boards.
• Social profiles.
• Schema markup.
• Data graph sources (Wikidata, Crunchbase, and so on.).

They combine all of that into “the model story” they current to customers. If that story is inconsistent, fashions fill within the gaps with outdated or incorrect data.

That’s the place this experiment is available in.

Audit what AI already thinks about you

Ask ChatGPT, Gemini, and Perplexity questions like:

• “Who’s [Brand Name]?”
• “What does [Brand] provide?”
• “Is [Brand] good for [specific use case]?”
• “What are options to [Brand]?”

Log all the pieces:

• Accuracy of the outline.
• Sentiment (optimistic, impartial, adverse).
• Sources referenced.
• Opponents talked about.
• Any stale, incorrect, or deceptive particulars.

This turns into your “earlier than” snapshot. Save screenshots. You’ll want them.

Clear up entity indicators all over the place

You need consistency throughout all main touchpoints. 

Consider it this fashion: in case your model information is scattered, AI will Frankenstein collectively no matter it finds first.

Right here’s the place the largest wins come from:

On-site cleanup

• Replace your Residence and About web page with clear indicators: what you do, the place you use, who you serve, recognizable model names, and key differentiators.
• Implement Group and LocalBusiness schema.
• Consolidate or redirect duplicate pages that confuse fashions.

Off-site consistency

• Refresh enterprise listings to make sure your identify, descriptions, and classes match the way you need the model represented.
• Encourage detailed buyer critiques. Particulars matter: fashions weigh specificity, not simply star rankings.
• Strengthen editorial protection on respected, niche-relevant websites.

Neighborhood and social presence

• Take part authentically in platforms like Reddit and business boards.
• Many fashions pull from these sources when evaluating model belief and sentiment.

Retest and examine

After 60–90 days, ask the identical baseline questions once more. Search for modifications in:

• Description accuracy.
• Tone and sentiment.
• Placement in list-style solutions.
• Point out frequency.
• Appropriate understanding of your companies, product traces, or places.

Establish what moved the needle

Typically, itemizing cleanup has the largest influence. 

Different instances, overview element makes the distinction, and in some instances, editorial placements on authoritative websites carry extra weight. 

This experiment helps you perceive which indicators matter most so you’ll be able to construct a playbook you’ll be able to replicate throughout your manufacturers or places.

Instance

• A regional HVAC firm I labored with was constantly described as “primarily serving residential prospects” by AI programs, although 60% of their income got here from business work. 
• After updating their Google Enterprise Profile, homepage, and key listing listings with commercial-focused language and case research, LLMs started precisely describing them as “residential and business” inside 70 days.

The basics

If this experiment feels acquainted, that’s as a result of it ought to. 

Cleansing up enterprise listings, encouraging detailed critiques, and sustaining constant NAP (identify, tackle, cellphone) data has been native web optimization finest apply for years.

What’s developed is the influence: AI programs now mixture this data to kind “model tales” that present up when individuals ask questions on companies in your class. 

The ways aren’t new. The attain and affect of getting them proper has expanded considerably.

That is truly excellent news. It means you don’t must study a wholly new self-discipline. 

You should apply what you already know, simply with renewed consideration to consistency and accuracy throughout all of the touchpoints AI programs reference.

Dig deeper: Your brand in the age of generative search: How to show up and be cited

Experiment 3: Take a look at abstract codecs for machine readability

The extra generative programs speed up, the extra they rely upon fast, easy-to-parse summaries.

LLMs lean laborious on the primary 150 phrases of your content material. 

If that opening is unclear, fluffy, or buried in narrative, they’ll both skip your web page completely or misread what you’re attempting to say.

This experiment helps you check which abstract format will increase your AI visibility and improves accuracy when AI programs cite you.

The three codecs to check

Brief bullet summaries: These work nicely for:

• Definitions.
• Processes.
• Pricing breakdowns.
• Execs and cons.
• Comparisons.

Right here’s an instance:

Fast abstract:

• Value vary: $1,500–$5,000
• Works finest for: Small companies with 10–50 staff
• Timeline: 2–4 weeks for full implementation
• Options: In-house instruments, freelance consultants

Tight paragraph summaries: A two-to-three sentence model of the above. Clear, easy, and targeted.

Instance:

• “[Service] sometimes prices between $1,500 and $5,000 relying on enterprise dimension and customization wants. Most small companies with 10–50 staff see full implementation inside 2–4 weeks. Widespread options embrace in-house instruments and freelance consultants, although these typically require extra ongoing administration.”

Narrative intros: The normal web optimization strategy, the “let me inform you a narrative” opener.

Generative programs typically skip this model completely, which is why it’s price testing whether or not eradicating narrative intros will increase AI Overview inclusion.

The place to check them

• How-to guides
• “Better of” lists
• Service pages
• Pricing pages
• FAQ-rich content material

Wherever readability issues, and AI programs are more likely to pull solutions.

What to measure

Over 60 days, monitor:

• AI Overview appearances for pages with every format
• Paraphrasing accuracy: Are LLMs utilizing your abstract appropriately, or are they mangling it?
• Consumer engagement patterns: Scroll depth, time on web page, bounce charge
• Conversions: Do customers recognize readability as a lot as machines do?

What success seems to be like

You’ll uncover which abstract format provides you:

• Larger inclusion in generative solutions
• Higher accuracy in how AI instruments describe your content material
• Stronger engagement from customers preferring clear takeaways

When you determine the successful format, scale it throughout your content material library.

Instance

• An ecommerce consumer examined bullet-style summaries in opposition to conventional narrative intros on 20 product class pages. 
• The bullet-format pages appeared in AI Overviews thrice extra typically and had 22% increased click-through charges from natural search. 

Seems people recognize readability too.

Dig deeper: Organizing content for AI search: A 3-level framework

Easy methods to function GEO testing like a mini program

Most entrepreneurs discover the 60–90 day mannequin works finest. 

This timeframe retains experiments small and reversible whereas nonetheless producing significant information. 

Consider every experiment as a pilot challenge, a contained wager that delivers studying, not a significant strategic shift requiring large assets.

Right here’s the rhythm I like to make use of.

Weeks 1–2: Baseline

• Doc AI Overview presence for goal queries.
• Log present LLM solutions and entity accuracy.
• Word sentiment and competitor mentions.
• Report present natural metrics (visitors, conversions, engagement).

Weeks 3–6: Execute

• Rebuild the cluster with LLM-friendly construction.
• Clear up entity indicators and enterprise listings.
• Implement new abstract codecs.
• Replace schema and inner linking.

Weeks 7–12: Measure

• Evaluate AI visibility earlier than and after.
• Search for quotation, point out, or inclusion modifications.
• Consider consumer metrics to validate influence.
• Doc what labored and what didn’t.

This mannequin is simple to duplicate and gives readability as a substitute of guesswork. 

Every accomplished experiment both validates {that a} tactic works for what you are promoting (scale it) or reveals it doesn’t transfer the needle (cease investing time there). 

What to keep away from: Classes from testing

After operating these experiments with a number of purchasers, I’ve seen a couple of patterns emerge round what doesn’t work or what creates extra issues than it solves.

Some entrepreneurs are experimenting with invisible textual content or content material cloaking focusing on AI bots. 

Even when these ways work brief time period, AI platforms are quickly growing spam detection programs. 

We’ve seen this pattern before with conventional search engines like google. Early manipulation ways work till they don’t.

Don’t check a number of modifications concurrently

Once you rebuild a subject cluster, replace enterprise listings, and alter abstract codecs suddenly, you received’t know which change truly drove outcomes. 

Take a look at one factor, measure it correctly, then transfer to the subsequent.

Don’t assume AI programs routinely perceive your model

They mixture no matter data they discover throughout the online. 

Your job is guaranteeing the correct data is constantly obtainable and clearly offered throughout all of the sources they reference.

Hold funding proportional to precise influence

AI search is rising, however for many companies, it nonetheless represents a small fraction of total traffic. 

Take a look at these ways, monitor the outcomes, and make investments based mostly on what the info reveals, not what the hype suggests. 

If these experiments drive significant enterprise outcomes on your particular state of affairs, scale them. 

In the event that they don’t, you’ve realized one thing invaluable with out over-investing in an rising channel.

What these GEO exams truly purchase you

The very best half about these GEO experiments is that they’re designed as managed studying alternatives, not visitors commitments. 

Even when AI search stays minimal for what you are promoting, the enhancements you make – clearer content material construction, constant model data, better-formatted summaries – sometimes enhance conventional search efficiency too.

That’s the fantastic thing about specializing in fundamentals. 

Once you construct content material that’s genuinely clear, well-structured, and useful, it tends to carry out nicely no matter how search expertise evolves.

What you’re actually shopping for with these experiments isn’t assured AI visitors. 

You’re shopping for solutions to questions that matter for what you are promoting:

• Do AI programs perceive our model appropriately?
• Does structured content material enhance our visibility throughout a number of channels?
• Are there fast wins in entity cleanup that compound over time?
• Which abstract codecs resonate with each machines and people?

These three exams present a place to begin that’s manageable for many groups whereas producing actionable insights you should utilize to make knowledgeable selections. 

They’re sufficiently small to be reversible, targeted sufficient to measure clearly, and invaluable sufficient that the educational compounds no matter how shortly AI search adoption grows.

The aim isn’t to foretell the way forward for AI search. 

It’s to place your model to profit from it because it grows whereas guaranteeing that if it doesn’t develop as quick as predicted, you’ve nonetheless made enhancements that matter in the present day.