We make B2B brands quotable by AI search engines
Generative Engine Optimization for companies whose buyers now ask ChatGPT, Perplexity and Gemini instead of scrolling ten blue links.
Most GEO advice stops at "add schema and write better content". Ours starts with two questions we can answer with data: is anything actually crawling your site, and what does an engine have to quote if it wanted to cite you. We publish the audit criteria, the weights, and our own numbers — including the ones that are unflattering.
- 12audit checks, weights published
- 60terms defined, no hedging
- 33crawler signatures tracked
- 0fabricated case numbers
Definition
What GEO actually is
One paragraph, then the table that matters.
| Dimension | SEO | GEO |
|---|---|---|
| Competition unit | A page versus a page, ranked in a list | A passage versus a passage, merged into one answer |
| Delivery mechanism | Index → rank → click | Retrieve → extract → generate → cite (click optional) |
| Content unit | Page with a target keyword | Self-contained claim with entity, quantity, time and scope |
| Primary signal | Links, relevance, intent match | Crawl access, extractability, factual density, entity consistency |
| Metric | Position, impressions, clicks | Mention rate, citation rate, answer accuracy |
| Time to effect | Weeks to months | Retrieval: hours. Model memory: months. |
| Failure mode | Ranking on page 3 | Being quoted wrong, or not quoted at all |
Note on naming. You will also see AEO (Answer Engine Optimization) and LLMO. They describe overlapping work. We use GEO because it names the mechanism — a generative engine — rather than the interface. The Chinese version of this site has the full methodology.
Delivery
Four workstreams, in this order
Sequence matters. Fixing content before crawl access is wasted effort.
Crawl and index access
Confirm the right bots are allowed and the wrong ones are blocked. This means distinguishing training crawlers (GPTBot, ClaudeBot, CCBot) from retrieval crawlers (OAI-SearchBot, PerplexityBot) — a distinction most robots.txt files get wrong.
- Per-engine robots.txt review, not a blanket allow
- Sitemap hygiene: only indexable URLs, no tag-archive bloat
- Render check: what an engine sees without JavaScript
Content an engine can lift
A model quotes sentences, not pages. We rewrite so the quotable sentence exists, is self-contained, and survives being taken out of context — with the number, the period and the scope attached to it.
- Claim-level structure: answer first, evidence second
- Factual density: specific quantities instead of adjectives
- FAQ blocks written as questions buyers actually type
Machine-readable markup
JSON-LD that matches what is visible on the page. Mismatched markup is worse than none — it teaches the engine that your structured data is unreliable.
- Organization, Service, Product, FAQPage, BreadcrumbList
- Entity definitions consistent across every page
- Syntax and validity checked, not just presence
Honest measurement
We tell you which numbers are real and which are structurally unavailable. Conversation-style engines do not return citation URLs, so their citation rate is zero by design — reporting it as a KPI would be dishonest.
- Mention rate and answer accuracy, tracked separately
- Server-log verification of crawler traffic
- Explicit "not measurable" list in every report
Measurement
The 12 checks we score, with the weights
Same weights as our platform. No hidden scoring, no proprietary black box.
The on-page audit runs twelve checks totalling 100 points. We publish the weights because a score you cannot reproduce is a sales device, not a measurement.
| Check | Weight | What it catches |
|---|---|---|
| Body content | 16 | Too thin to be worth extracting |
| Meta title & description | 10 | No clear claim in the snippet |
| Internal links | 10 | Orphan pages an engine cannot reach |
| CITE coverage | 10 | Context, intent, trust, evidence gaps |
| Heading structure | 8 | Claims buried under decorative headings |
| Schema markup | 8 | Missing or mismatched JSON-LD |
| Images | 8 | Missing alt text, oversized assets |
| FAQ block | 7 | No question-shaped content |
| Factual density | 7 | Adjectives where numbers belong |
| H1 | 6 | Missing, duplicated or off-topic |
| Canonical | 5 | Duplicate URL signals |
| H2 structure | 5 | Flat hierarchy, no scannable sections |
What we measure
- Whether a brand is mentioned at all, per engine, per question
- Whether the mention is accurate (price, scope, capability)
- Which passage was cited, when a citation URL exists
- Whether the page was fetched, verified in server logs
What nobody can measure
- Citation counts for engines that return no source URLs
- AI traffic split inside Google Search Console — it is not broken out
- Which training run your content entered, or whether it was retained
- Future answer positions — they are generated per query, not indexed
Evidence
We ran the numbers on ourselves first
Four days of raw server logs, published with the method and the caveats.
Across 111,731 requests over four days, 22.3% came from crawlers identifying as AI. That headline number turned out to be nearly useless — and finding out why was more valuable than the number itself.
The last one is worth stating plainly, because we maintain an llms.txt file ourselves: in this window, not one of the major AI crawlers requested it. We keep publishing it because it costs little, but we do not sell it as a result.
Engagement
Four ways to work with us
Pick by how much internal capacity you have, not by budget alone.
| Model | Best when | What you get |
|---|---|---|
| Diagnostic | You suspect a problem but do not know where | Full audit, crawler log analysis, prioritized fix list. Fixed scope, fixed fee. |
| Retainer | You have a site and a team, but no GEO capability | Monthly audit, content production, technical fixes, measurement report. |
| Embedded (FDE) | You need the capability inside your own workflow | An engineer working with your team on your backlog, with your tools. |
| Outsourced | You have no in-house content or technical resource | End-to-end execution: research, writing, publishing, monitoring. |
Fit
Who this works for, and who it does not
Saying no early saves both sides a quarter.
A good fit
- B2B companies with a real product and buyers who research before contacting
- Sites with at least a few dozen pages of substantive content, or budget to build it
- Teams willing to give publishing access and review drafts within a week
- Companies that want to see the raw data, not just a dashboard score
A poor fit
- Anyone who needs a guaranteed ranking or citation count in a contract
- Pure affiliate or thin-content sites — GEO cannot fix a site with nothing to say
- Projects that require us to publish claims the client cannot substantiate
- Buyers who want a fixed number of articles and nothing else
FAQ
Questions buyers actually ask
Is GEO just SEO with a new name?
No, though the overlap is real. SEO competes for a position in a list; GEO competes for a passage inside a generated answer. The skills overlap — crawl access, structure, clarity — but the unit of work is different: a self-contained claim rather than a page targeting a keyword. Pages that rank well often contain nothing an engine can quote cleanly.
How long before we see anything?
Retrieval changes can show up within days — if a bot is blocked today, unblocking it is fast. Model memory is slower: content incorporated into training data affects answers months later, on the model vendor's schedule, not ours. We separate these two timelines in every report because conflating them is how agencies end up promising results they cannot control.
Can you guarantee we will be cited by ChatGPT?
No, and treat anyone who does as a red flag. Citation depends on the engine's retrieval index, the specific query, the competitive set for that query, and the model version in use that day. What we can commit to is the work: verifiable technical fixes, published content, and measurement you can audit.
What do you need from us to start?
Three things: read access to your analytics (Google Search Console if you have it), publishing or staging access to the site, and one decision-maker who can approve drafts. Diagnostic engagements also benefit from 30 days of raw server logs — that is where the crawler evidence lives.
Do you work in English, Chinese, or both?
Both. Our team operates in Chinese and English, and we have delivered content for B2B brands selling into English-language markets as well as Chinese-language GEO programmes. Content is written in the target language by a writer fluent in it — we do not machine-translate and call it localization.
How do you handle the data we give you?
Under a data processing agreement that covers roles, sub-processors, cross-border transfer and a 48-hour breach notification. If you need a signed copy or need to complete your own vendor security questionnaire, we will do that before the engagement starts — not after.
Why publish your own bad numbers?
Because 86% of our crawl budget going to tag archives is a useful finding, and hiding it would make every other number we publish less credible. Our own site is the one case study a reader can verify line by line. If the method does not work on us, it will not work on you.
Start with the audit, not the proposal
Tell us the domain. We will run the 12 checks and send you the scored result — including the parts that fail — before any discussion of scope or fee.
中文版:www.semkw.com · Team and roles: how we are organized