AI Is Starting to Make Buying Decisions: What GEO Fieldwork Across 10 Industry Brands Reveals

Gravity industry insight and GEO fieldwork white paper | 10 brand client groups · 11 delivery scopes · Retesting across ChatGPT, Gemini, and Perplexity

When consumers stop searching and start asking, a brand's first competitive battleground changes.

“Which modular energy-storage system should I buy for 72 hours of home backup?” “How should a hybrid-work team choose ergonomic chairs in bulk?” “What fuel-line specification does an E85 build need?” More buying decisions now take place inside an answer from ChatGPT, Gemini, or Perplexity. The pages AI cites and the criteria it uses to frame its advice can strongly influence whose recommendation the buyer follows.

That is the problem GEO—generative engine optimization—addresses. Yet many GEO offerings still focus on publishing content at scale, while leaving three questions that brands actually care about unanswered:

• Has AI actually come to read my website?
• After reading it, does AI cite and adopt the content?
• How can we demonstrate these changes to executives and clients?

Over the past several months, Gravity has turned these questions into verifiable evidence across 10 brand client groups and 11 delivery scopes. Here is what we observed.

1. The Results: Six Sets of Numbers

Measured results through late October 2026 are shown below. Brand names are anonymized by industry.

In one sentence: AI-engine visits increased substantially, including the effect of expanded monitoring coverage, while AI answers began frequently citing official brand pages and adopting their product-selection criteria.

2. Why the Approach Works: One Business Thread and Four Verifiable Evidence Layers

Many GEO services concentrate on generic traffic-oriented articles, overlooking a basic feature of AI answer generation: AI does not lack encyclopedic concepts. It lacks product facts and comparison boundaries that answer specific buying constraints.

Gravity's core methodology: Start with consumers' real buying questions; connect product facts, official-site content, external evidence, and AI answers; identify the gaps affecting recommendations; and improve them through implementation and retesting.

This business thread helps us avoid three common pitfalls:

2.1 Replace “Keywords” with Real Buying Tasks to Find Recommendation Gaps

Consumers do not enter isolated keywords into AI. They ask decision questions with constraints. Gravity turns each industry's demand into standardized buying-task question clusters and uses paired T0-baseline/T1-retest comparisons on the same questions to locate gaps that obstruct recommendations:

• Brand- and model-anchored questions: does AI recognize the brand, and does it assign core specifications and prices to the correct product?
• Controlled rephrasings: when a buyer asks differently, is the official target page still cited consistently?
• Negative-boundary questions: when asked “Is it right for everyone?” or “Does installation require no electrician?”, can AI make an accurate conditional recommendation?
• Unbranded natural-discovery questions: can the brand enter the shortlist when buyers do not name it, or explicitly seek an alternative to a competitor?

We report “being named” and “being proactively recommended” separately. A high mention rate on branded questions measures the ability to support existing brand interest; unbranded recommendations measure a breakthrough in open-market discovery.

2.2 Four Evidence Layers, Each with a Traceable Source

• Read: did 17 categories of dedicated AI engines—OpenAI, Anthropic, Perplexity, Google, Meta, and others—fetch pages, and how did crawl volume change? A subset has additional network-signature verification: 3,345 of the latest 23,483 visits. The rest are self-declared through User-Agent strings.
• Cited: did the final answer from ChatGPT, Gemini, or Perplexity display an exact link to a brand's official page? Unused back-end candidates do not count.
• Adopted: did the answer reproduce the decision criteria, engineering examples, or audience boundaries in the modules we added?
• Captured: did AI referrals reach the site and convert? This layer is connected client by client through authorized GA4 or Shopify data. The data in this article stop at the first three layers.

3. Separating Signal from Noise: Fewer Total Bots Does Not Mean Less AI Interest

Clients often ask: “Why did our total Bot count fall this week?”

For North American off-grid solar brand B, total sitewide Bots fell from 75,040 to 47,045. On the surface, that looks like a decline. Reconciliation showed a different picture:

• Ad inspections receded: Meta ad-landing-page inspections alone decreased by 22,597 events, explaining about 80% of the overall decline.
• AI-classified events grew: the five AI event categories rose from 2,436 to 8,101 (+232.6%). Model crawling grew +923.7%, and OAI-SearchBot increased from 49 to 410.

The same distinction held for NAS brand C and wool-apparel brand I: broad sitewide Bot totals fell, while core dedicated AI engines grew +308.93% and +1,980.00%, respectively.

For brand teams: do not let one total Bot number lead the analysis. Understanding who visits and what they read is the first lesson of GEO. Visits outside the 17 core categories are not necessarily noise; we simply analyze them separately from clearly classifiable dedicated AI engines.

4. Four Industry Cases: Four Ways to Solve Different Commercial Problems

From projects that have completed retesting, we identified four representative problem-solving paths. Each links to a full standalone case.

Case 1 | Home-Backup and Whole-Home Energy-Storage Brand A: From Supporting Brand Interest to Unbranded and Competitor-Alternative Recommendations

• Commercial challenge: in a home-backup market with purchases costing thousands of dollars, buyers ask for “Powerwall alternatives,” small-capacity systems that expand over time, or air-conditioning backup through multi-day hurricane outages. The brand had zero recommendations on some important unbranded baseline questions. Branded runtime and new-product questions also exposed a lack of standard engineering examples incorporating loss factors.
• Gravity's work: mapped 22 buyer Demand Cells; launched a Home-Backup Runtime Estimation Guide with the complete 12.288 kWh nominal → 8.51 kWh usable → approximately 20.8-hour engineering example, a modular-expansion guide, and competitor-alternative comparison content; then completed 585 answers across 65 questions and 93 fact checks.
• Observed change: on a strictly unbranded “Powerwall alternatives” question, Perplexity repeatedly cited the new comparison page and placed the brand in “The strongest Tesla alternatives to compare” leading shortlist. Gemini's official-site citation rate on natural questions more than doubled: 14.29% → 38.64% average, with a 48.48% peak. Recommendations broke through from zero on hurricane-backup and 5–30 kWh expansion questions. On the original 11 diagnostic questions, official-site citation share rose by about 20 percentage points, 43.12% → 63.56%. The runtime guide received 4 direct citations, and its 20.8-hour calculation was reproduced at near-verbatim level.
• Read more: Home-backup and whole-home energy-storage GEO case →

Case 2 | Smart Ergonomic Office-Furniture Brand D: Turning Complex Model Differences into AI-Usable Buying Criteria

• Commercial challenge: differences among Standard, SE, Gen, and Pro models were complex, without a clear official decision table. AI could mix up battery runtime, lumbar-tracking motors, and accessory compatibility.
• Gravity's work: planned and helped deliver 4 new or improved structured pages addressing key buying decisions: Four Flagship Models Compared, Hybrid-Team Pilot Guide, Battery Runtime FAQ, and Accessory Compatibility Self-Check. All are now public.
• Observed change: core dedicated AI Bot visits increased 66 → 791 (+1,098.48%). All 4/4 target pages were cited during retesting, generating 20 exact-page citations across 18 answers. The model-comparison page alone contributed 14/20 citations: Gemini 10 and Perplexity 4.
• Read more: Smart ergonomic office-furniture GEO case →

Case 3 | North American Automotive-Performance Brand E: Preserving the Sales First Screen while Adding Engineering Guidance

• Commercial challenge: enthusiasts need fuel-compatibility, AN-size, and fuel-sender matching guidance, but product pages held only basic size tables. AI often gave generic recommendations based on CFM alone. Meanwhile, the high-converting first-screen shopping area could not be casually changed.
• Gravity's work: used non-intrusive additions—preserving the entire first-screen purchase area and adding 19 engineering-selection paragraphs in the detail sections of the carburetor collection, PTFE fuel-line, and fuel-tank pages.
• Observed change: live readback confirmed 19/19 paragraphs implemented verbatim, versus 0/19 in historical pre-change snapshots. The 3 pages received 8 exact citations: carburetors 5, across ChatGPT, Gemini, and Perplexity; PTFE fuel line 2 from Gemini; fuel tank 1 from Perplexity. Answers directly adopted the engineering principle that carburetor selection must consider engine use, secondary actuation, and fuel pressure rather than CFM alone.
• Read more: North American automotive-performance GEO case →

Case 4 | Global Fashion-Gifting DTC Brand F: Improving Recommendations under Complex Conditions, Not Just Exposure

• Commercial challenge: the category is crowded, buyers often omit brand names, and ask difficult boundary questions such as whether every product suits everyone.
• Gravity's work: launched a gifting-scenarios guide; fact-checked 342 candidate brand statements extracted from AI answers; and proposed one-sentence, minimal wording alignment, pending client approval, while preserving core benefits such as 30-day returns and exchanges.
• Observed change: core AI Bot visits increased 377 → 12,414 (+3,192.84%). ChatGPT cited the gifting page and used it to structure recommendations. On the relevant 36 ChatGPT negative-control questions, conditional recommendations increased 9/36 → 19/36 (25.00% → 52.78%). Of 342 candidate statements, 170 were supported by official-site evidence, 151 awaited verification, and 21 were not assessed; there were 0 factual conflicts within the checked scope.
• Read more: Global fashion-gifting DTC GEO case →

5. Client Coverage: 10 Industries, from High-Ticket Hardware to Tourism

Brands are at different stages: some have completed multiple retest samples, some have just launched infrastructure, and others are building a baseline. The business thread is consistent: find the fact and content gaps affecting AI recommendations, then help AI read, cite, and use the information correctly.

6. Three Recommendations for Brand Decision-Makers

• Open the machine-access channel before scaling content. If firewalls or robots rules unintentionally block dedicated AI engines, or key specifications are buried in scripts and images, more articles cannot solve the underlying access problem. Cross-checking server logs against dashboard data is the first step.
• Give AI a decision table with boundaries, not more adjectives. Models need structured answers bringing together scenarios, objective limits, currently sold SKUs, and transparent return rules on a single page. Clearer information makes adoption easier.
• Report an evidence chain, not a single number. Connect implementation, official-site readback, AI crawling, exact citations, and adoption in answers. Each step needs a source. That is how GEO withstands executive scrutiny.

7. Measurement Notes

Long-term trust starts with clear boundaries:

• Time windows: crawl comparisons use each brand's first project-period 7 days versus the latest 7 days, October 17–23, 2026. October in the monthly curves includes only October 1–23.
• Crawl scope: only visits declaring one of 17 dedicated AI products in the User-Agent are counted; ad inspections, social-app embedded browsers, and dual-purpose general crawlers are excluded. Only a subset has additional network-signature verification: 3,345 of the latest 23,483. The 7.5× figure includes expanded monitoring coverage, with 14,042 latest-period visits from server-side layers absent initially; it must not all be interpreted as net growth in actual visits. Monthly curves use a broader “AI-classified events” scope and cannot be added directly to dedicated-engine counts.
• Citation scope: only exact target-page links displayed in final answers count. Chart 3's 34 citations / 32 answers / 9 pages covers targeted-page retests for furniture, automotive, and NAS brands: 25 brand/model-validation citations, 6 controlled-comparison citations, and 3 category-shopping citations. Furniture pages were new or improved; automotive pages received additive modules. Evidence attributing every page's final published version is not yet fully closed.
• Home-backup retest scope: 65 standard questions comprise 46 original questions and 19 client-added questions. The original 46 had a T0 baseline of 138 answers, 46 × 3 platforms. T1 covered all 65 across 3 platforms and 3 independent same-day rounds: 65 × 3 × 3 = 585 answers, 580 with verifiable reference links. Paired baseline comparisons apply to the original-question group; the added 19 are evaluated separately and do not all have an old baseline.
• Conditional-recommendation scope: 25.00% → 52.78% comes from 36 ChatGPT negative-control questions, not the aggregate recommendation rate across 3 platforms.
• Misses are recorded too: in the same retest cycle, the NAS two-factor-authentication guide received 0/15 citations, while a solar brand's article published one week earlier had not yet been cited. These observations inform the next optimization round.
• Causality: these changes were observed during retesting alongside infrastructure and content launches. They are not strict causal attribution or a promise of future behavior by any AI platform.

Where Does Your Brand Stand in AI Answers?

Apply the same questions to your category: when buyers ask ChatGPT, Gemini, and Perplexity, whose pages does AI cite, and how does it describe your brand?

Contact the Gravity team to explore a GEO implementation path for your industry.

About Gravity

Gravity is a technology service provider focused on generative engine optimization and AI-native growth. We start with consumers' real buying questions; connect product facts, official-site content, external evidence, and AI answers; identify gaps affecting recommendations; and improve them through implementation and retesting—helping brands get read, cited, and adopted in ChatGPT, Perplexity, Gemini, and leading local-market AI models. Our work spans consumer electronics, clean-energy storage, smart offices and homes, automotive performance, fashion, and cross-border tourism.