GEO basics you can do yourself
Ten complete guides, every step written so you can follow it. The first three cost nothing, don't need us and can be finished today, so try them yourself first, then decide whether you need help.
Ten guides, two tracks
The first three are free foundational steps. The next seven get into execution detail.
Start here · free foundational steps
- 015 min How to test your brand's AI visibility in 5 minutes 5 steps
- 0230 min How to turn TDS specs into AI-readable structured data 5 steps
- 0310 min How to configure robots.txt to let major AI crawlers in 5 steps
Go further · execution detail
- 0420 min How to write an llms.txt (and whether it does anything yet) 7 steps
- 05Half a day How to build a competitor-model mapping library 7 steps
- 0610 min per article How to tell whether AI will cite a piece of technical content 7 steps
- 071 hr How to structure FAQs so AI can extract them easily 7 steps
- 081 hr per month How to track your brand's AI mention rate over time 7 steps
- 092 hrs How to choose industry platforms worth publishing on 7 steps
- 103 hrs per article How to write technical long-form that convinces international buyers 8 steps
- Not sure where to start? Take the 18-item GEO self-check Start the self-check
How to test your brand's AI visibility in 5 minutes
No tools and no budget. In five minutes you'll know whether AI puts you on its shortlist when buyers ask.
Write down 5 questions your buyers actually ask
Write them as a buyer, not as a seller. For example: "Who are reliable suppliers of thermostable feed enzymes?" or "Who supplies bearings equivalent to SKF 6205-2RS?" Don't type your company or product name. Buyers don't know your name yet, and that is exactly the problem you're testing.
Ask each question in 4 AI models
Run every question in ChatGPT, Perplexity, Claude and Gemini. Ask in the language your buyers use, which for most export categories is English. Start a new chat each time so earlier context doesn't skew the answer.
Record three things
Which companies the AI names, whether your brand appears, and in what position. Also note which source links the AI cites.
Follow up: "Are there alternatives to the big names?"
This step tells you the most. If the AI only now comes up with a few names, and reluctantly, challenger suppliers as a group have thin source content in your category. That is your window of opportunity.
Screenshot everything as your baseline
This is the starting point for measuring results. Re-run the exact same questions in three months. The difference is your result.
How to turn TDS specs into AI-readable structured data
AI misreads specs locked in PDFs. Move them into HTML and add Schema markup, and AI can cite them with confidence.
Pick the specs that actually drive selection
Don't move everything. Pick the handful of specs buyers use to filter suppliers. For chemical raw materials that's usually specific surface area, activity units, particle size distribution and purity. For coating additives: matting efficiency, clarity and surface feel. For precision parts: dimensional tolerance, surface roughness and hardness. The rule is simple: start with the numbers buyers ask for first when they request a quote.
Build one HTML page per product
One grade, one page. Use the full grade name as the title, and list specs, units and test methods in a table. Keep the PDF as a downloadable attachment, but don't let the PDF be the only version.
Add Product structured data
In JSON-LD, mark up name, sku and manufacturer, and put the specs in an additionalProperty array, each with name / value / unitCode. Always use UN/CEFACT codes for units (e.g., M2/G, ML/G).
Add applications and comparable models
Use applicationCategory to state where the product is used, and isSimilarTo to point to international models with comparable specs. State spec facts only. Don't assert "equivalent" or "replacement".
Validate with official tools
Run the page through Google's Rich Results Test and the Schema.org Validator to make sure there are no syntax errors. If the markup has errors, AI ignores the whole block.
How to configure robots.txt to let major AI crawlers in
Many manufacturer websites block AI crawlers by default and nobody on the team knows. This is the first gate in GEO.
Check your current status
Open yourdomain.com/robots.txt in a browser. If you see Disallow: / or a rule blocking GPTBot, AI can't read any of your content right now.
Confirm the list of crawlers to allow
At minimum: GPTBot, OAI-SearchBot, ChatGPT-User (OpenAI); ClaudeBot, Claude-User, Claude-SearchBot (Anthropic); PerplexityBot, Perplexity-User; Google-Extended; Applebot-Extended.
Write an explicit Allow for each one
Give every User-agent its own block with Allow: /. Don't rely on User-agent: * as a catch-all. Some crawlers only follow the block that names them.
Declare your sitemap
Add Sitemap: https://yourdomain.com/sitemap.xml at the end of the file so crawlers know where to start.
Re-check after it goes live
Reload yourdomain.com/robots.txt to confirm the change is live. Then check whether your server or CDN has its own user-agent blocking rules. That layer is often the real culprit.
How to write an llms.txt (and whether it does anything yet)
The short answer: write one, but don't expect it to work on its own. Here's why, and how.
Understand where it actually stands
llms.txt is a community convention, not a W3C or IETF standard. As of 2026, none of the major vendors (OpenAI, Google, Anthropic and others) has publicly committed to reading it in production systems, and in our tests AI crawlers rarely fetch the file. Almost all of them still crawl HTML directly. Anyone who tells you "add an llms.txt and AI will recommend you" is wrong.
So why write one at all
Two real reasons. First, AI coding agents in IDEs (Cursor, Claude Code, GitHub Copilot and others) routinely fetch /llms.txt and /llms-full.txt, and some of your technical customers use these tools. Second, writing the file forces you to turn "who we are, what we sell, what the specs are, why you can trust us" into one unambiguous statement of fact. You can reuse that statement on your website, external platforms and product pages, which is where it actually does its work. The cost is close to zero. Just don't treat it as your main lever.
Put it in the root directory, written in Markdown
The path is always https://yourdomain.com/llms.txt. Start with an H1 containing your company name, followed immediately by a one-sentence positioning statement in > blockquote format. That sentence must stand on its own, because it's the part most likely to be lifted whole.
Lock down your numbers in a "Key facts" section
Brand name (and any local-language name), scope of products or services, target markets, ranges for core technical specs, certifications and standards, contact details. One item per line, each starting with a hyphen. Every number needs a unit and a test method. These facts must match your website and external platforms exactly. When they don't, AI can't cross-verify them and simply uses none of them.
Make the "Pages" section a link list with descriptions
Format: - [Page name](full URL): what this page covers. Don't just list links. The one-line description is what the model reads. List only pages with real information; skip pages like "Contact us".
Add an llms-full.txt if you need one
llms.txt is the index; llms-full.txt is the complete fact file, with definitions, methods, specs and positions written out in full. That's how we run our own site, so you can compare against our /llms.txt and /llms-full.txt directly.
Finally, do the thing that matters more
Copy the positioning statement and key facts from llms.txt into the body text of your homepage, your About page and your profiles on external platforms. The text in your HTML pages is what crawlers actually read. llms.txt is insurance, not the main lever.
How to build a competitor-model mapping library
One of the most common questions international buyers ask AI is "What's an alternative to model X?" Without a mapping library, you'll never be in that answer.
Pick the models to map against, no more than 10
Choose the incumbent models your main products actually compete with for orders. Base it on real feedback from your sales team: whichever brand's datasheet customers bring to compare prices against is the one to map. Doing too many dilutes quality. Ten done thoroughly beat fifty done superficially.
Use only publicly available datasheets as your basis
The competitor's TDS on their own website, public product catalogs, industry database entries. Never use internal documents that circulate privately, or competitor reports a customer gave you. If the source can't be cited publicly, the whole comparison falls apart, and it can lead to disputes.
Build a spec-by-spec table with three columns
Column one: spec name and test method. Column two: the competitor's published value. Column three: your measured value. Don't hide the differences. State them plainly: "Ours: XX. Theirs: YY." AI actually trusts content that admits differences more, because that's what real technical documentation looks like.
Write a one-sentence conclusion for each mapping
Something like: "In [specific application], the [key spec] of [your grade] falls in the same range as [competitor model], so it can be evaluated as an alternative." The sentence must stand on its own out of context, because AI usually extracts just one sentence.
Keep the wording on the right side of compliance
Use only "spec comparison", "can be evaluated as an alternative" and "in the same range". Never use "equivalent", "identical", "official replacement" or "same as OEM". Use other companies' trademarks descriptively only, never in your product name or as the lead of a title. This isn't excessive caution: misleading comparative claims are actionable in the US and the EU.
Mark up the relationship with Schema
In the product page's JSON-LD, use isSimilarTo to point to the mapped model and put the specs in additionalProperty. That way machines read the relationship with certainty instead of guessing from the body text.
Publish one page per model; don't combine them
Give each mapping its own page, with a title in the buyer's own question format, e.g., "Alternatives to [model]: Technical Comparison". One big all-in-one comparison article ends up matching none of the queries.
How to tell whether AI will cite a piece of technical content
Run these seven checks before you publish. Publishing only what passes all seven works far better than publishing ten pieces and hoping.
Is there a conclusion sentence that can be extracted on its own?
AI extracts sentences, not paragraphs. Every section needs at least one conclusion sentence that stands on its own out of context and contains a specific subject and number. "This product delivers excellent performance" doesn't count. "Tensile strength 45 MPa (ISO 527-2), suitable for injection-molded automotive interior parts" does.
Do the numbers have units and test methods?
AI won't cite a bare number, because it can't tell what basis it was measured on. Every value gets a unit, and key specs get the test standard number. This is usually the check with the lowest pass rate.
Does the information match what's published elsewhere?
AI prefers to cite content that multiple independent sources can cross-verify. The same spec must be identical on your website, on external platforms and in your product catalog. Three places with three different numbers equals zero usable places.
Does it answer a question someone actually asks?
If the title isn't a sentence a buyer would type, the content will most likely never be triggered by any query. Use real phrasings like "How to choose...", "What is the difference between..." and "Alternatives to...", not "Introduction to Product XX".
Is the structure machine-friendly?
Correct heading hierarchy, specs in HTML tables rather than images, FAQPage or HowTo markup, one idea per paragraph. A spec table embedded in an image might as well not exist.
Is it clear who is speaking and whose product it is?
A machine reading the page must be able to tell who is making the claims and whose product they're about. Use the full brand name and grade name. Don't refer to everything as "we" or "this product" throughout.
Does it overlap with what you've already published?
If the title structure, paragraph skeleton and opening sentences look like content already on your site, it gets classified as low-quality mass-produced content. Better to publish two pieces this week than ten that look alike.
How to structure FAQs so AI can extract them easily
FAQs are the highest-return GEO content format: the question is the query and the answer is the unit AI extracts. But most companies write them wrong.
Questions must be the buyer's own words, not your sales pitch
Wrong: "What are the advantages of our products?" No buyer has ever asked that. Right: "Can this grade be used in waterborne systems?" "What's the minimum order quantity?" "Is it REACH-registered?" Where do the questions come from? Copy them straight from questions customers have actually asked in sales emails.
Give the complete answer in the first sentence
AI usually extracts only the first sentence of an answer, so that sentence must be the full conclusion, with no lead-in. "Yes. This grade is REACH-registered, registration number XXXX." Then add the details. Opening with "That depends on several factors..." gives up your chance of being extracted.
One question, one answer; don't pack five sub-questions into one
Combined questions produce extracted answers that no longer match the original question. Split them into five entries instead.
Include specific numbers and conditions in the answer
"Wide temperature range" becomes "Operating temperature -20 °C to 180 °C; continuous use recommended at no more than 150 °C." An answer without numbers adds nothing when AI cites it, so AI doesn't cite it.
Add FAQPage structured data
Use JSON-LD to mark up each Question / acceptedAnswer pair. The text in the markup must match the visible text on the page exactly. If it doesn't, it's treated as manipulation and the markup for the whole page is ignored.
Write each language version natively; don't machine-translate
Your buyers ask in their own language, usually English, and stiff machine translation significantly lowers the odds of being cited. In each language version, phrase the questions the way buyers in that market actually say them, not as a literal translation of your source-language version.
Revisit after launch
After three months, check which questions actually earned AI citations and replace the ones that didn't. An FAQ is a living document, not a one-time deliverable.
How to track your brand's AI mention rate over time
Without a baseline there's no result to measure. This method needs no paid tools. A spreadsheet is enough.
Fix one set of questions and never change it
12–20 questions covering four types: general category (who can supply X), spec filter (which X products withstand temperatures above 150 °C), alternatives (alternatives to model Y) and regional (X suppliers in a given country or region). Once the set is fixed, don't change it. If you do, you lose comparability.
Start a new chat for every test
Chat history seriously contaminates results. Same model, same question, new chat every time, and make sure no logged-in personal memory is active. Two tests run under different conditions produce meaningless numbers.
Record four fields
Whether the brand is named (yes/no), its position (which name it was in the list), which source links the AI cites, and whether the AI describes you accurately. The third field is the one people most often skip, yet it tells you exactly where to put content next.
Screenshot every test
AI answers can't be reproduced, so without a screenshot you have no evidence. Keep timestamped screenshots and file them as date-model-question number.
Re-test on a fixed date every month
For example, run one round on the 1st of every month: all 4 models × all questions. Once a month is enough. Test more often and most of what you see is noise.
Calculate three numbers
AI mention rate = times named ÷ total questions asked. Average recommendation position = the mean position when named. Source hit rate = times the AI cited your own assets ÷ times named. The third number tells you whether your content is actually doing the work.
Test your competitors alongside you
Your own numbers mean little without a reference point. Record how 2–3 main competitors perform on the same question set. If your mention rate rose 10% while a competitor's rose 30%, you actually lost ground.
How to choose industry platforms worth publishing on
External source content is what keeps your mention rate stable. Pick the wrong platforms, and the money and time are wasted.
First check whether the platform lets AI crawlers in
Open platformdomain.com/robots.txt and see whether it blocks GPTBot, ClaudeBot or PerplexityBot. If it does, however much authority the platform has, AI can't read what you publish there. This is an automatic disqualifier, yet almost nobody checks.
Test instead of guessing
Ask ChatGPT and Perplexity the typical questions in your category and see which sites they actually cite. The platforms AI already cites are the ones worth publishing on. That's more reliable than any domain-authority ranking.
Expect very different source preferences across models
Large-scale citation analysis from 2026 shows that the domains cited by ChatGPT and Perplexity overlap by only about 11%. The two run on almost entirely separate sets of sources. So don't expect one platform to cover everything; cover at least two or three different kinds of sources.
Don't underestimate community platforms
The same data shows community platforms (Reddit, Quora and similar) account for about 52.5% of citations, more than brands' own domains at 47.5%. A genuine technical discussion in a professional community can be more likely to get cited than a polished long-form article on your website. That only works with real participation, not ads.
Check that content stays accessible long term
Requirements: each piece has its own permanent URL, can be viewed without logging in, and won't be archived or taken down by the platform a few months later. Content behind a login wall is worth almost nothing for GEO.
Confirm what the platform allows you to do
Read the platform's content rules line by line: whether you can mention your brand, include outbound links, or publish product comparisons. Breaking the rules gets posts deleted at best and accounts banned at worst, and everything you invested goes to zero.
Tier your effort; don't spread it evenly
Choose 2–3 high-value platforms that AI already cites for in-depth technical content, and 3–5 broad-reach platforms for consistent baseline information. The first group builds credibility; the second builds cross-verification.
How to write technical long-form that convinces international buyers
When a supplier's technical article fails to convince international buyers, the problem is rarely grammar. It's structure and calibration.
Make the title a sentence a buyer would type
Not "Company X's High Quality Products". Write "How to Select Sealing Materials for High-Temperature Hydraulic Systems" or "Alternatives to [model]: A Parameter-Level Comparison". The title decides whether the content gets triggered by any query at all.
State the core conclusion within the first 100 words
Technical readers check the conclusion first and then decide whether to read the details. AI extracts the same way. Put the "so what" up front. Don't build up layer by layer and reveal the point at the end.
Let data do the talking; cut every adjective
Words like "excellent performance", "superior quality" and "world-leading" say nothing in a technical context and can even reduce credibility. Replace all of them with specific values and test methods. This usually cuts a third of the length and makes the article stronger.
Admit limitations; it raises credibility
A statement like "Not recommended for systems with pH < 4" convinces technical buyers more than an article full of strengths, and it reads like real technical documentation. An article that only says good things reads like an ad.
Break it up with subheadings, one idea per paragraph
The readability of long-form content depends heavily on subheadings. Use H2/H3 to split the article into 5–8 sections, and open each with one sentence saying what that section covers. This also makes it easier for AI to locate and extract.
Put every spec in a table, with units and standards
Use HTML tables, not images. Write units in standard international notation (m²/g, not m2/g), and give the exact ISO, ASTM or national standard number for each test method.
Get a native-speaker technical review, not just language polishing
What matters isn't whether the grammar is correct. It's whether the terminology is right and the tone fits industry norms in your target market. The best reviewer is a native-speaking technical professional in the industry; the next best is a senior sales rep with years of selling the category internationally.
Run the citation self-check before publishing
Go through the seven checks in "How to tell whether AI will cite a piece of technical content" above. If it doesn't pass, don't publish.
Do the three basics yourself, then see how big the gap still is
Let AI crawlers in, put your specs in HTML, write an llms.txt. Those three steps usually fix "invisible to AI". What's left, competitor-model mapping and external source content, is the part we specialize in.