Generative engine optimization: what AI says about you, and what 91 industrial websites reveal
Generative engine optimization (GEO) is the practice of shaping how AI answer tools such as ChatGPT, Perplexity and Google's AI Overviews find, describe and cite a business. Search engine optimization aims for a ranking on a list of links. GEO aims for something different: a correct mention inside the answer itself. The term was introduced in a 2023 research paper by Aggarwal and colleagues, later accepted to the KDD 2024 conference.[1]
This essay brings together three bodies of evidence. The first is behavioural data on how people react when an AI answer appears. The second is experimental data on which kinds of content those answers draw on. The third is our own scan of 91 US mid-cap manufacturing and supply-chain companies, which shows how ready a typical industrial website is to be read by machines. We close with a framework, the Retrievability Ladder, and a step-by-step audit any marketing team can run in an afternoon.
Idea in brief
The problem. When Google shows an AI summary, people click a regular search result in 8% of visits, against 15% when no summary appears.[2] The answer now does more of the selling. If it leaves you out, or gets you wrong, the buyer may never reach your site.
Why it happens. AI answers lean on sources that are easy to read, consistent across the web, and full of checkable evidence. In controlled tests, adding quotations, statistics and cited sources raised a page's visibility in AI answers by up to about 40%. Keyword stuffing made it worse.[1]
The gap. In our scan of 91 industrial companies, the basics were mostly in place: 91% of readable homepages linked to an about page and 87% had a meta description. The machine layer was thin. Only 49% carried Organization schema, 12% published an llms.txt file, and 26% of homepages did not return a readable page to a research script that named itself.
The answer. Work up the Retrievability Ladder in order: be reachable, be readable, be consistent, be corroborated by other sites, then be quotable. Facts come before files.
Key findings
The essay in 2 minutes- Finding 115% → 8%
AI summaries roughly halve clicks to regular results.
In Pew Research Center's March 2025 browsing panel, the click rate fell from 15% to 8% of visits when an AI summary appeared. Links inside the summary were clicked in 1% of visits.[2]
- Finding 260%
Buyer-style questions trigger AI answers most.
Searches that began with who, what, when or why produced an AI summary 60% of the time, against 8% for searches of one or two words.[2]
- Finding 327.2
Evidence beats keywords.
Of nine content edits tested on the GEO benchmark, adding quotations scored highest (27.2 against a baseline of 19.3). Keyword stuffing scored below the baseline (17.7).[1]
- Finding 4+115%
Smaller players gain most.
For sources ranked fifth in search, citing sources lifted AI-answer visibility by 115%. For sources ranked first, the same edit cut it by 30%.[1]
- Finding 51%
Industrial websites are half-built for machines.
Among 67 readable homepages in our sample, 57% had any structured data and 1% had FAQ schema. Across all 91 companies, 12% published an llms.txt file.
- Finding 67%
Few industrial firms block AI crawlers on purpose.
Of 70 companies with a readable robots.txt file, 5 (7%) blocked at least one AI crawler, and none blocked the crawlers OpenAI uses for search and for user-requested visits.
Chapter 1 · 1 min
What generative engine optimization is
GEO sits beside SEO rather than replacing it. Google states that there are no extra technical requirements to appear in AI Overviews or AI Mode beyond being indexed and eligible to show a snippet.[3] So a site that is invisible to search will usually be invisible to AI answers too. What GEO adds is attention to how a page reads once it is lifted out of context and placed inside a generated answer.
Terms used in this essay
| Term | Meaning |
|---|---|
| Generative engine | An AI system that writes an answer from several sources, such as ChatGPT search, Perplexity or Google AI Overviews. |
| GEO | Generative engine optimization: improving how often, and how accurately, a source appears in those answers. |
| AI visibility | How often a business is mentioned in AI answers to the questions its buyers ask. |
| Structured data (schema) | Code that labels facts on a page, such as an organization's name, logo and address, using the schema.org vocabulary. |
| llms.txt | A proposed Markdown file at a site's root that summarises the site for AI tools. Published as a proposal in September 2024, not a standard.[4] |
| Retrievable brand | A brand whose facts are easy for both people and machines to find, check and repeat correctly. See the Hive entry on what a retrievable brand is. |
GEO and SEO compared
| Search engine optimization | Generative engine optimization | |
|---|---|---|
| Goal | Rank on a results page | Be mentioned and described correctly in an answer |
| What the buyer sees | A list of links | A written answer, sometimes with sources |
| What matters most | Relevance, links, page quality | Consistent facts, outside mentions, quotable evidence |
| How it is measured | Rankings and clicks | What AI tools say when buyers ask |
| Overlap | Large. Good SEO is the foundation for GEO. |
Chapter 2 · 1 min
The evidence, part 1: what happens when an AI answer appears
Pew Research Center tracked the real browsing of 900 US adults in March 2025 and published the results in July 2025.[2] It is one of the few studies of AI answers based on observed behaviour rather than surveys.
Share of Google search visits (%), 0 to 20
Figure 1. Share of Google search visits that ended in a click. Source: Pew Research Center, 2025.
Three numbers stand out. People clicked a regular result in 15% of visits without an AI summary and 8% with one. They clicked a link inside the summary in 1% of visits. And they ended their browsing session on 26% of pages with an AI summary, against 16% of pages with regular results only.[2] For a business, that means the summary is often the last thing a buyer reads.
The second finding matters more for B2B marketers, because it shows which searches get an AI answer at all.
Share of searches with an AI summary (%), 0 to 70
Figure 2. Share of Google searches that produced an AI summary, by type of search. Source: Pew Research Center, 2025.
Across all searches in the study, 18% produced an AI summary. Short searches of one or two words rarely did (8%). Searches of ten words or more did 53% of the time, and searches starting with who, what, when or why did 60% of the time.[2] Those long, question-shaped searches are exactly how a buyer describes a real problem: "Who can machine aluminium housings for medical devices in small batches?" That is where AI answers concentrate, and where a supplier is either named or left out.
Tip: Pew counted one context only, Google search in the US. Treat the figures as direction for your own market, not a forecast.
Chapter 3 · 2 min
The evidence, part 2: what makes a source quotable
The original GEO paper tested how different edits to a web page changed its visibility in AI-written answers.[1] The team built GEO-bench, a benchmark of 10,000 queries from many domains. They ran each query through a generative engine built on GPT-3.5 using the top five Google results as sources, then rewrote those sources in nine ways and measured how much of each source made it into the answer.
Visibility score (position-adjusted word count), 0 to 30
Figure 3. Visibility of a source after each edit, measured as position-adjusted word count. Source: Aggarwal et al., KDD 2024, Table 1.
The three edits that add evidence did best: quotations (27.2), statistics (25.2) and cited sources (24.6), against an unedited baseline of 19.3. Keyword stuffing, the oldest SEO habit, scored 17.7, below doing nothing. The authors report that the best methods can raise visibility by up to about 40%, and that the effect held when they repeated the test on Perplexity, with gains of up to 37% there.[1]
The finding with the biggest commercial meaning is in the paper's second table.
Change in AI-answer visibility (%)
Figure 4. Change in AI-answer visibility by the source's original search rank. Source: Aggarwal et al., KDD 2024, Table 2.
For a source ranked fifth in search, citing sources raised its visibility by 115%, adding quotations by 100% and adding statistics by 98%. For the source already ranked first, the same edits lowered visibility by 21% to 30%.[1] Evidence is a lever the challenger can pull. A mid-sized manufacturer does not need to outrank the category leader to become the source an answer leans on.
Limits of the experimental evidence
- The study measured how much of a source appeared in the answer, not clicks, leads or revenue.
- The main engine was built on GPT-3.5. Commercial AI tools have changed since, and they do not publish how they pick sources.
- The authors found that results varied by domain, so the best edit for a legal query may differ from the best edit for an engineering one.
Read the paper as strong direction, not as a formula.
If you want a team to run this kind of evidence review on your own pages, the Brave New audit checks how your business reads to buyers and to the machines that now answer for them.
Chapter 4 · 3 min
Original research: how ready are industrial websites for AI?
The studies above describe AI answers in general. We wanted to know where industrial companies, the sector we work in most, stand today. So on 3 November 2026 we scanned the public websites of 91 US mid-cap manufacturing and supply-chain companies. The sample is every company in the S&P MidCap 400 and S&P SmallCap 600 lists classed under eleven manufacturing and freight sub-industries, from industrial machinery and building products to aerospace, trucking and marine transport. The full method is at the end of this essay.
Finding 1: the basics are in place, the machine layer is not
Share of companies (%), 0 to 100
Homepage signals n=67; sitemap and llms.txt n=91.
Figure 5. Share of companies showing each signal. Homepage signals are based on the 67 homepages that returned a readable page; sitemap and llms.txt are based on all 91 companies. Source: Brave New Media scan, 3 November 2026.
The signals any web agency checks were mostly present. Among the 67 homepages we could read, 91% linked to an about or company page, 87% had a meta description and 82% had an H1 headline. Across all 91 companies, 74% published a sitemap.
The signals that help machines identify a company were much thinner. Only 57% of readable homepages carried any structured data, and 49% carried Organization schema, the markup that tells a machine the company's official name, logo and profiles elsewhere. One homepage in 67 had FAQ schema. And 11 of 91 companies (12%) published an llms.txt file, which is still only a proposal.[4]
This is a gap, not a scandal. Google says no special schema is needed to appear in its AI features.[3] But Organization schema is the cheapest way to tell every machine, in one place, which facts about a company are official. Half the sample has not done it.
Finding 2: one in four homepages turned away a declared research bot
Number of companies, of 91
Figure 6. What 91 industrial homepages returned to a scripted request that identified itself as a research script. Source: Brave New Media scan, 3 November 2026.
Our script named itself openly as a research bot, as AI crawlers do. 67 of 91 homepages (74%) returned a readable page. 17 (19%) refused the request outright with a 403 error, and 7 more returned some other response, such as a redirect or a rate limit.
This does not prove those sites block AI crawlers. Firewalls often treat known crawlers from verified addresses differently from an unknown script. It does show that many industrial sites sit behind security rules that nobody in marketing has reviewed with AI crawlers in mind. Ask your web team one question: what does our firewall do when OAI-SearchBot, PerplexityBot or ClaudeBot asks for our homepage?
Finding 3: few companies block AI crawlers on purpose
70 companies had a robots.txt file we could read. Five of them (7%) blocked at least one AI crawler. Four blocked GPTBot, four blocked Google-Extended, four blocked Common Crawl's CCBot, three blocked ClaudeBot and two blocked PerplexityBot. None blocked OAI-SearchBot or ChatGPT-User.
That split is worth understanding. OpenAI documents GPTBot as the crawler that collects content for model training, OAI-SearchBot as the one that surfaces sites in ChatGPT search, and ChatGPT-User as the agent that fetches a page when a user asks for it.[5] A company can decline training and still appear in search answers. The five companies in our sample that block AI crawlers appear to have made that kind of deliberate choice. The larger risk is the accidental one in Finding 2.
Tip: Open yoursite.com/robots.txt today. If you see "Disallow: /" under a user agent you do not recognise, look it up before you change anything. Blocking a training crawler and blocking a search crawler have different costs.
Chapter 5 · 1 min
A framework: the Retrievability Ladder
The research points to an order of work. Each rung depends on the one below it. A quotable page does nothing if the crawler cannot reach it, and schema does little if the facts it labels contradict your LinkedIn page.
| Rung | Question it answers | Evidence | What to check |
|---|---|---|---|
| 5. Quotable | Is there evidence worth lifting into an answer? | Quotations, statistics and cited sources gave the largest lifts in the GEO tests.[1] | Specific numbers, named sources and short self-contained answers on key pages |
| 4. Corroborated | Do other sites confirm what you say? | AI tools build answers from several sources; 88% of Google's AI summaries cited three or more.[2] | Reviews, directories, trade press, partner and association pages |
| 3. Consistent | Do your facts match everywhere? | Machines cross-check sources, and disagreement forces a guess. | Name, services, markets, location and contact details across every profile |
| 2. Readable | Can a machine tell who you are? | 49% of our industrial sample had Organization schema. | Organization schema, clear about page, sitemap, server-rendered text |
| 1. Reachable | Can a crawler get to the page at all? | 26% of homepages in our sample did not return a readable page to a declared bot. | Firewall rules, robots.txt, crawler access by user agent |
Most teams start at the top, rewriting copy to sound quotable, because that feels like marketing. The data suggests starting at the bottom. In our sample, the cheapest fixes (firewall review and Organization schema) were the ones most often missing.
Chapter 6 · 2 min
How to audit what AI says about your business
An AI visibility audit starts with buyers, not with your website. If you only test questions written from your own pages, you only learn whether AI repeats your pages. You learn nothing about whether you show up when a buyer describes a problem in their own words.
Step 1: Collect real buyer questions
Pull questions from the places buyers speak for themselves: sales call notes, RFQ emails, support tickets and the questions prospects ask at trade shows. Write ten, in three kinds:
- Category questions: "Who makes custom corrugated packaging for frozen food in the Midwest?" These matter most, because the buyer has not heard of you yet.
- Comparison questions: "What should I look for in a contract manufacturer for medical parts?"
- Name questions: "What does [your company] do, and is it any good?"
Step 2: Ask three tools, in fresh sessions
Run each question in ChatGPT, Perplexity and Google (read the AI Overview or AI Mode answer). Start a new chat for each question so earlier answers do not steer later ones.
Step 3: Log four things per answer
Use one row per question and tool. Record whether you were mentioned, whether the description was correct, which competitors appeared, and which sources were cited. The cited sources are the most useful column. They show you which sites the AI trusts in your category, and they become your target list for rung 4.
Step 4: Repeat by market and by quarter
If you sell in more than one country, run the same questions in each market's language. Answers that are right in English can be wrong or empty in German or Polish, and the cited sources change from market to market. Then repeat every quarter. AI answers shift, and a dated log is the only way to see whether your fixes worked.
By hand this is slow, so most companies check once and stop. It is the kind of job we build AI flows for: the same question set run across tools and markets on a schedule, with changes flagged for a person to review. One small team can then watch ten markets as easily as one.
Chapter 7 · 1 min
How to fix what AI says about you
Work up the ladder. Each fix below maps to a rung.
1. Reachable. Ask your web team to test the homepage and two key product pages with the main AI crawler user agents, and to review firewall and bot-management rules. Decide on purpose which crawlers to allow, using each vendor's documentation.[5]
2. Readable. Add Organization schema with your legal name, logo, address and sameAs links to your official profiles. Make sure your key pages show their main text in the HTML, not only after scripts run. The structured data should match the visible page.[3]
3. Consistent. Make your name, services, markets, location and contact details read the same on your website, Google Business Profile, LinkedIn, directories and review sites. The Hive explains why brand consistency matters for AI search.
4. Corroborated. Get onto the sites your audit showed AI already cites for your category: trade directories, industry associations, trade press and partner pages. The Hive's guide on how to get mentioned in ChatGPT and Perplexity covers this step by step, and how AI models choose brands explains why some names keep appearing.
5. Quotable. Rewrite key pages so each section answers one buyer question in its first two sentences, then supports it with a real number, a named source or a short quote. Add FAQ sections built from your audit questions. The Brave New Q&A shows the format.
An llms.txt file is cheap, so add one if your web team has time, but expect little. It is a proposal, not a standard, and Google says no special AI text files are needed for its AI features.[3][4] The Hive's entry on what llms.txt is covers the format.
Which fixes to do first
| Fix | Ladder rung | Effort | Likely impact | When |
|---|---|---|---|---|
| Firewall and robots.txt review | Reachable | Low | High if you are blocked | This week |
| Organization schema | Readable | Low | Medium | This month |
| Consistent facts everywhere | Consistent | Low to medium | High | This month |
| Mentions on cited sites | Corroborated | Medium to high | High | Ongoing |
| Evidence-rich answers on key pages | Quotable | Medium | High | As pages are updated |
| llms.txt | Readable | Low | Low and uncertain | Any time |
Chapter 8 · 1 min
What this means for leaders
For a manufacturing or supply-chain business, AI answers change the economics of being found. A long, specific buyer question used to return ten links, and the best-optimised site took the click. Now it often returns one written answer, and the click rate on everything else falls.[2] The suppliers named in that answer get the first call.
The research also suggests the change favours challengers that bring evidence. A mid-sized supplier with clear facts, consistent profiles, outside mentions and real numbers can be the source the answer leans on, even against larger rivals that rank higher in classic search.[1] That makes GEO less a technical project than a growth one: it is how a company with a strong product gets onto shortlists in markets where nobody knows its name yet.
If you want an AI flow that checks what AI says about you in every market you sell to, and flags when it changes, start with the Brave New audit.
Chapter 9 · 1 min
Common mistakes
- Testing only your own words. Prompts generated from your website tell you whether AI repeats your website, not whether buyers find you.
- Starting at the top of the ladder. Quotable copy cannot help a page a crawler cannot reach.
- Keyword stuffing. It scored below the unedited baseline in the GEO tests.[1]
- Blocking crawlers by accident. Security rules set years ago can turn away the crawlers that feed AI search.
- Treating files as a fix. Schema and llms.txt help machines read facts. They cannot invent facts that are missing.
- Inventing numbers to look quotable. An AI repeating a false claim about you is worse than silence.
Chapter 10 · 1 min
Methodology
Sample. All 91 companies listed on Wikipedia's S&P MidCap 400 and S&P SmallCap 600 constituent pages under the GICS Industrials sector in eleven sub-industries: industrial machinery and components; building products; aerospace and defense; cargo ground transportation; construction machinery and heavy transportation equipment; electrical components and equipment; trading companies and distributors; agricultural and farm machinery; air freight and logistics; marine transportation; and industrial conglomerates. Official websites came from Wikidata (property P856) and were corrected by hand where Wikidata was missing or pointed to a regional site (five companies).
Scan. On 3 November 2026, a script requested each company's homepage, robots.txt, llms.txt and sitemap.xml. It identified itself with the user agent "BraveNewResearch/1.0" and a link to brave-new.com. It read raw HTML only and did not run JavaScript or attempt to get around any block.
Measures. A homepage counted as readable if it returned HTTP 200. Homepage signals (about link, meta description, H1, JSON-LD, Organization schema, FAQ schema) were measured on the 67 readable homepages. An llms.txt file counted if it returned plain text or Markdown rather than an HTML or error page; all 11 positives were checked by hand. A sitemap counted if robots.txt declared one or /sitemap.xml returned a valid sitemap. A crawler counted as blocked if robots.txt disallowed the whole site for that user agent, or for all agents with no specific group for it.
Limits. This is a snapshot of one day and one sample of US-listed companies. A firewall that refused our script may treat verified AI crawlers differently. Homepage checks say nothing about deeper pages. We did not measure what AI tools say about these companies.
The data behind every chart in this essay is published alongside it as CSV files.
Chapter 11 · 2 min
Frequently asked questions
What is generative engine optimization?
Generative engine optimization (GEO) is the practice of making AI tools such as ChatGPT, Perplexity and Google's AI answers find, describe and cite a business correctly. It overlaps heavily with SEO but aims for an accurate mention inside the answer rather than a ranking. The term comes from a 2023 research paper accepted to KDD 2024.
Does GEO actually work?
In controlled tests on 10,000 queries, adding quotations, statistics and cited sources raised a source's visibility in AI answers by up to about 40%, and by up to 37% on Perplexity. Sources ranked lower in search gained the most. The tests measured visibility in answers, not sales.
How do you get your business to show up in ChatGPT?
Make sure crawlers can reach your site, add Organization schema, and keep your facts the same across your website, Google profile, LinkedIn and directories. Earn mentions on the sites AI tools already cite for your category. Then write key pages with direct answers, real numbers and named sources.
How can you check what AI says about your business?
Collect ten real buyer questions from sales calls and emails, including category questions that do not use your name. Ask each one in ChatGPT, Perplexity and Google in fresh sessions. Record whether you are mentioned, whether the description is right, which competitors appear and which sources are cited. Repeat every quarter.
Should we block AI crawlers?
That depends on the crawler. OpenAI, for example, separates GPTBot, which collects training data, from OAI-SearchBot, which surfaces sites in ChatGPT search. Blocking training while allowing search is a common choice. In our scan of 91 industrial companies, 7% of those with a readable robots.txt blocked at least one AI crawler, and none blocked OpenAI's search crawler.
Can a marketing agency do generative engine optimization for my business?
Yes. An agency can run the buyer-question audit, review crawler access, fix inconsistent facts, add schema, win mentions on cited sites and rewrite key pages around evidence. Brave New Media is a marketing agency that does this work, with experience in supply chain and manufacturing, and builds AI flows that track what AI tools say about a business across markets.
Your next step
This week, open yoursite.com/robots.txt, ask your web team what your firewall does with AI crawlers, and type your three hardest buyer questions into ChatGPT, Perplexity and Google. Paste the answers into one document and mark every word that is wrong or missing. That list, sorted by the ladder, is your GEO plan.
If you want a team to run it across every market you sell to, start with the Brave New audit.
Brave New Media is a marketing agency that builds AI flows to help companies scale their ideas, grow their customer base and expand into new markets, with experience across supply chain and manufacturing.
See also
- What is generative engine optimization, Brave New Hive
- What is a retrievable brand, Brave New Hive
- How AI models choose brands, Brave New Hive
- What is llms.txt, Brave New Hive
- The Brave New Business Journal, weekly thinking on business, money and AI
References
1. Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K. and Deshpande, A. "GEO: Generative Engine Optimization." Proceedings of KDD 2024. arxiv.org/abs/2311.09735
2. Pew Research Center. "Google users are less likely to click on links when an AI summary appears in the results." 22 July 2025. pewresearch.org
3. Google Search Central. "AI features and your website." developers.google.com
4. Howard, J. "The /llms.txt file." Proposal, September 2024. llmstxt.org
5. OpenAI. "Overview of OpenAI crawlers." platform.openai.com/docs/bots
6. Brave New Media. AI-readiness scan of 91 US mid-cap industrial company websites, 3 November 2026. Data published with this essay.