Two GEO experiments challenge conventional AI visibility advice

Much of the existent proposal astir generative motor optimization (GEO) is based connected theory, isolated screenshots, aliases a azygous campaign. I wanted measurable results, truthful I ran 2 system experiments backmost to back, tracked the results manually, and documented what worked, what failed, and what changed betwixt the 2 tests.

The first experiment, for an existing marque we consulted on, costs thousands of dollars and ran complete respective months. The 2nd was a 30-day cold-start research a SaaS nexus building agency pinch nary measurable AI beingness erstwhile the trial began.

Each trial tracked 15 commercial-intent keywords. The first covered 4 AI platforms, while the 2nd covered six. I logged 775 citation events crossed some experiments, and 1 of my original conclusions didn’t past the 2nd test. Let’s excavation into really I group these up, what I observed, and really you tin use it to your work.

Experiment 1: The advisor test

We tracked 15 commercial-intent keywords crossed 4 platforms: ChatGPT, Claude, Gemini, and Perplexity. Every query was tally manually, pinch and without a VPN, to relationship for imaginable location-based differences.

Our strategy was to spot listicles connected sources that were commonly surfaced by LLMs for our keywords, arsenic good arsenic connected the brand’s site. We supported these done PR, impermanent posts, and integrated LinkedIn activity.

By the end, the marque appeared for astir 10 to 12 of the 15 keywords. Peak keyword beingness reached 37.01% connected April 29. The wide inclination was upward, pinch meaningful week-to-week variation. The citations by level were arsenic follows: ChatGPT 148, Claude 96, Gemini 87, and Perplexity 64.

The starring root for citations was a broad listicle connected Indeed SEO, pinch 190 mentions. MEXC and a GlobeNewswire merchandise followed, but neither came adjacent to the aforesaid volume. In all, listicles accounted for 72.4% of citations and PR 24.1%. Guest posts, the owned site, and LinkedIn divided what was left.

We besides learned 4 awesome lessons you tin use to your adjacent project.

Listicle placement, PR, and impermanent posts reenforce each other

In this experiment, listicles and PR appeared to activity arsenic 1 strategy alternatively than arsenic abstracted channels. The listicles that earned citations were usually the aforesaid ones being amplified done PR. The listicle introduced the claim, the property merchandise reinforced it, and impermanent posts referenced both. Each furniture appeared to execute amended erstwhile supported by the others.

We recovered securing a placement connected a root the exemplary already cites tin compound visibility quickly. Here, Indeed SEO appeared astir 10 times successful ChatGPT’s answers earlier we contacted the publication.

After our placement went live, it remained our largest citation root for months. That placement generated much citations than each the others combined.

Guest posts appeared to widen the effect of the PR placements. Claude, successful particular, cited impermanent posts that referenced features successful Yahoo and Business Insider, moreover erstwhile it didn’t mention those publications directly.

Source decay is accelerated and authority matters

Roughly half of the sources stopped being cited wrong 30 days. One MEXC placement fell from 29 mentions to 11 week complete week, while a Triple Review listicle that had been cited consistently declined without immoderate nonstop intervention. The consequence suggests that one-time publication is improbable to prolong visibility connected its own.

In our tests, root authority and relevance appeared to matter much than contented value alone. A applicable level was:

• Government and acquisition sources.
• News publications.
• Industry-relevant sites.
• General sites.

In Phase 1, manufacture placements supported by PR produced beardown results. In Phase 2, better-written listicles connected wide sites produced small visibility. PR appeared tin of amplifying a beardown placement, but not compensating for a anemic source.

Depth appeared to matter much than frequency. Our first 5 listicles covered the taxable only briefly, and astir generated small visibility. The 1 broad portion continued to gain citations.

The adjacent group appeared to power visibility. One listicle placed america alongside Lily Ray and Aleyda Solis. After those names were removed, capacity declined wrong days. This suggests that the models whitethorn measure the surrounding entities, not only the individual mention. A akin shape appeared successful PR, wherever being named alongside recognized experts outperformed a standalone characteristic connected a stronger outlet.

SERP visibility still appeared to power LLM visibility, particularly successful ChatGPT. When the exemplary relied connected web hunt to resoluteness a query, brands absent from the retrieved results were besides absent from the answer.

Capitalization and query type matter

The capitalization of queries appeared to impact which sources were retrieved. Capitalized and lowercase versions returned different citations successful 3 repeated tests, though this uncovering requires further validation.

Query type appeared to power the root types selected by the models. Software and instrumentality queries favored high-authority reappraisal sites, while work queries much often returned listicles. Matching the placement type to the query appeared to amended the likelihood of being cited.

Exact-match keywords and reply placement are important

Exact-match keyword targeting still appeared to matter. We classed for “Best LLM SEO Consultant” but hardly appeared for “Best AI SEO Consultant,” contempt the akin intent. The first building had a dedicated listicle, while the 2nd didn’t. In this test, broader semantic sum didn’t span the gap, suggesting that high-value commercialized keywords whitethorn require dedicated assets.

In our tests, pages performed amended erstwhile the reply appeared wrong the first 100 words. A cardinal takeaway artifact adjacent the apical of the page produced a larger betterment than immoderate different on-page alteration we made.

Other contented findings include:

• FAQ contented performed amended erstwhile it was visible by default alternatively than hidden down expandable sections.
• Self-contained sections appeared to execute better. For example, “What to look for erstwhile hiring an LLM SEO expert” and “Where to prosecute one” worked amended arsenic abstracted sections than arsenic a azygous mixed section.
• Question-based headings besides performed amended successful our tests. For example, “How is AI SEO different from accepted SEO?” outperformed “AI SEO vs. accepted SEO.”
• Freshness besides appeared to matter. Recent data, existent references, and visible publication dates were associated pinch stronger citation performance.

We learned a batch from this first test.

The original scheme for information 2 was a database of things to test: Person schema, LinkedIn cadence, and a YouTube push. Instead, I sewage the chance to tally the full playbook from zero connected a different marque successful a different vertical, which is simply a acold amended trial of whether immoderate of the findings are generalizable.

Dig deeper: 3 GEO experiments you should effort this year

Experiment 2: The acold commencement test

The trial progressive a SaaS nexus building agency pinch nary measurable AI beingness erstwhile the research began. During the baseline model from April 30 to May 29, the marque had nary measurable beingness connected immoderate tracked platform. 

The research ran from May 30 to June 28, and covered six platforms, adding Google AI Mode and Grok to the database from the first experiment. We tracked 15 commercial-intent keywords that a SaaS purchaser mightiness usage while evaluating an agency. Every level and keyword was checked manually.

The research resulted successful 298 appearances successful 30 days from a opinionated start. The appearances by level were: Gemini 104, Google AI Mode 95, Claude 59, ChatGPT 32, Grok 4, and Perplexity 4.

Gemini and Google AI Mode accounted for astir two-thirds of the level totals listed above. This differed sharply from research one, wherever ChatGPT led. The comparison should beryllium treated cautiously because AI Mode wasn’t tracked successful the first experiment, and the 2 niches weren’t straight comparable.

Business impact: During the 30-day window, 18.5% of caller users arrived done referral traffic, and different 3.25% done GA4’s AI Assistant channel. Together, these channels accounted for conscionable complete one-fifth of each caller users. One Perplexity referral led to a imaginable who later became a paying customer, moreover though Perplexity was the lowest-volume level successful the experiment.

Here’s what this research taught me.

Focus connected observed citations and earned placements

I learned you want to build the target database from observed citations, not from DR alone. Before opening outreach, I ran each 15 keywords done each platform, logged the sources that appeared, and classed them by citation frequency. That classed database became the outreach list.

Five of 8 targets appeared successful the last citation mix. Indie Hackers accrued from 44 mentions during prospecting to 146 aft our placement went live, a 232% improvement. Bruce Jones SEO accrued from 26 to 69, up 165%, while TechBullion roseate from 15 to 37, a 147% jump.

The attack didn’t activity successful each case. RankTracker and HR.com showed less citations aft placement than during prospecting, and 2 shortlisted sites hadn’t appeared astatine each by the extremity of the measurement window. The method appeared to amended the deed rate, but it didn’t guarantee citations.

Placement attraction was extreme. Three sources — Indie Hackers, Bruce Jones SEO, and our ain listicle — accounted for 342 of 437 full root mentions, aliases astir 78%. Seven different unrecorded placements shared the remaining mentions. In this experiment, a mini number of sources drove astir of the visibility.

Within this experiment, earned placements outperformed owned contented by a wide margin. Of the aforesaid 437 root mentions, third-party listicles generated 85.8%, our self-published listicle generated 14.0%, and PR generated 0.2%.

Measure citations complete time

Time to citation ranged from 1 to 18 days. Two placements were cited the time aft publication, while others took 10 aliases 11 days. Four placements were unrecorded but hadn’t been cited by the extremity of the measurement period. The variety suggests that checking only once, 1 week aft publication, isn’t a reliable measurement approach.

The slowest root to beryllium cited was the 1 we controlled. Our ain listicle took 18 days, longer than each third-party placement, including 2 that were picked up overnight.

Owned listicles are a foundation, not a maturation lever

This is wherever I had to revise my original conclusion.

After research one, I recommended publishing listicles connected an owned tract because galore top-ranking brands appeared to use from their ain content. Experiment 2 suggested that an owned listicle is much of a instauration than a superior maturation engine. It was the slowest root to beryllium cited and contributed 14% of mentions, while earned placements generated astir of the visibility.

Dig deeper: How to cognize if your GEO is working

Comparative contented is effective

The owned listicle improved erstwhile it became little promotional and much comparative. On June 23, we updated it to see our starring competitors alternatively of presenting the marque alone. Mentions of that root roseate from 4 to 49, a 12.25-fold summation successful the last count. The regular visibility curve besides accrued during the aforesaid period, from 22 connected June 23 to a highest of 95 connected June 27.

This mirrored the peer-set effect observed successful research one, but from the other direction. Removing recognized names was followed by a diminution successful information one, while adding recognized competitors was followed by a important summation successful information two. The aforesaid shape appeared crossed 2 brands and 2 verticals, making it 1 of the findings I’d prioritize for further testing.

Comparative sum outperformed defense successful this test. The 12.25-fold summation followed an update that made the page much useful arsenic a class assets alternatively than arsenic a page focused chiefly connected our ain company. The domain, author, and keyword target remained the same. The main alteration was the scope of the content.

Citations don’t adjacent clicks

The most-cited and most-clicked sources weren’t the same.

Indie Hackers generated much citation measurement than immoderate different source, but its referral postulation remained flat. TechBullion produced less citations but accrued sessions from 1 to 64. Claude.ai referral sessions tripled, while ChatGPT referral sessions accrued by 166%.

These results propose that citation measurement unsocial isn’t capable for deciding which sources merit further investment.

Intent varies by model

In this experiment, Claude concentrated much heavy connected high-intent terms. Gemini led successful full appearances, 104 to 59, but Claude led aliases tied for first connected 5 of the 10 best-performing keywords.

Gemini’s measurement was distributed much broadly, while Claude’s was much concentrated connected commercialized queries. For a business evaluating level value, that distribution whitethorn matter much than the header total.

Exact-phrase assets besides performed well

Our 2 strongest keywords were “Best SaaS Link Building Agency successful USA” and the aforesaid building pinch “2026” added, pinch 18 appearances each. This repeated the shape from research 1 successful a different niche and suggests that dedicated assets whitethorn still beryllium basal for high-value commercialized keywords.

What I’d show personification starting today

These experiments provided maine pinch a coagulated database of lessons that tin thief you prioritize your efforts.

• Run your target keywords done the applicable AI platforms earlier investing successful placements. The sources already being cited should pass the outreach list, and that database whitethorn disagree importantly from a accepted prospecting spreadsheet.
• Budget for maintenance, not only first publication. In our first experiment, astir half of the sources stopped being cited wrong 30 days.
• Pay attraction to entity associations. Across some experiments, capacity changed erstwhile recognized companies aliases experts were added to aliases removed from listicles.
• Don’t dainty citation measurement arsenic the last business outcome. The astir often cited sources weren’t ever the strongest referral sources, and 1 low-volume Perplexity referral resulted successful a paying customer.

Traditional SEO focuses mostly connected ranking pages. Both experiments indicated that LLM visibility depends much heavy connected associations: which sources mention you, who appears alongside you, and really precocious those relationships were reinforced. The main disagreement was the domiciled of owned content, which the 2nd research showed was little powerful than I initially believed.

LLM visibility continues to evolve, but it’s ne'er excessively early to build connected these experiments and spot what your information tells you.

Dig deeper: GEO for group who person to deed gross targets