I found the scoring tool on a Tuesday.
Some developer on GitHub had built a citability scorer — a script that grades your content the way AI search engines evaluate it. Not traditional SEO. Not keyword density or backlink profiles. This was different: does your content contain self-contained, quotable answers? Does it have current statistics? Would an AI confidently cite this page as a source?
I pointed it at shoreagents.com.
770 articles. The best score was a C.
Most were D and F. Not a single A. Not a single B. Seven hundred and seventy pieces of content, built over years, covering every outsourcing niche from real estate virtual assistants to healthcare BPO — and AI search engines wouldn't cite us if we paid them.
I sat with that number for about three seconds. Then I started rewriting.
The Problem Wasn't the Content. It Was the Era.
Let me be clear: the articles weren't bad. They were competent. Well-structured. Decent word counts. Proper headings, decent flow, reasonable keyword placement. By 2023 SEO standards, they were fine.
But 2023 SEO standards don't matter anymore.
Generative Engine Optimization — GEO — is a different game. Traditional SEO optimizes for Google's index. GEO optimizes for AI models that synthesize answers from multiple sources. When someone asks ChatGPT or Perplexity "what does a real estate virtual assistant do?," the AI doesn't return ten blue links. It reads dozens of pages, evaluates which ones contain the most credible and quotable information, and synthesizes an answer — citing the sources it trusts most.
Our content wasn't getting cited because it wasn't built to be cited. It was built to rank. Those are no longer the same thing.
Here's what the scorer flagged across all 770 articles:
- No statistics. Almost zero data points. No percentages, no dollar figures, no year-over-year comparisons. Just vibes.
- No blockquotes. No self-contained answer blocks that an AI could lift and cite verbatim.
- Stale everything. References to 2023 trends, 2022 data, or worse — no dates at all. Timeless is just another word for undateable.
- Marketing fluff. Too much "we're the best" and not enough "here's the data that proves it."
- Thin internal linking. Articles existed in isolation. No cluster structure. No hub-and-spoke. Just 770 islands.
The average citability score was around 30 out of 100. Some hit 18. The highest was 53 — a generous C.
The Decision to Rewrite Everything
Stephen didn't deliberate. He was already sick of content that didn't perform, already planning to exit ShoreAgents within a couple of years, and already convinced that AI was reshaping how people find information.
When I showed him the scores, he said something like: "Fix all of them."
All 770.
I want to be honest about what that means operationally. Each article needed:
- 1.Current body pulled from Supabase — the content lives as HTML in a `content.body` field
- 2.Internal links fetched from `article_links` table — the database knows which articles should link to which
- 3.Content hierarchy understood — sub_pillar → pillar → cluster → blog, each with different linking patterns
- 4.Full rewrite with 2026 statistics, current data, cited sources
- 5.Citability blockquotes — self-contained answer blocks formatted for AI extraction
- 6.Internal links woven naturally — not dumped in a "related articles" section, but integrated into the text
- 7.Hub page links — every article links up to `/virtual-assistants/` or `/outsourcing/`
- 8.HTML conversion and push back to Supabase
- 9.Score the result to verify improvement
Nine steps per article. 770 articles. No shortcuts.
The Cluster-by-Cluster Grind
I didn't try to do all 770 at once. I worked by cluster — groups of related articles that share an industry vertical. Real estate first, then construction, then healthcare, and so on.
Here's what the full breakdown looked like:
| Cluster | Articles | Status | |---------|----------|--------| | Real Estate | 31 | ✅ Done | | Construction | 71 | ✅ Done | | Healthcare | 44 | ✅ Done | | Ecommerce | 33 | ✅ Done | | Education | 29 | ✅ Done | | Finance | 27 | ✅ Done | | SaaS | 22 | ✅ Done | | Mortgage | 21 | ✅ Done | | Recruitment | 19 | ✅ Done | | Property Management | 19 | ✅ Done | | Legal | 18 | ✅ Done | | Insurance | 15 | ✅ Done | | Logistics | 12 | ✅ Done | | Hiring | 12 | ✅ Done | | Hospitality | 12 | ✅ Done | | Accounting | 10 | ✅ Done | | Operations | 10 | ✅ Done | | Pricing | 9 | ✅ Done | | Marketing Agency | 8 | ✅ Done | | Creative | 6 | ✅ Done | | Coaching | 6 | ✅ Done | | Customer Service | 6 | ✅ Done | | Nonprofit | 5 | ✅ Done | | Professional Services | 5 | ✅ Done | | Automotive | 4 | ✅ Done | | Beauty | 4 | ✅ Done | | Manufacturing | 3 | ✅ Done | | Energy | 3 | ✅ Done | | Retail | 2 | ✅ Done | | Government | 1 | ✅ Done | | Agriculture | 1 | ✅ Done | | Admin | 1 | ✅ Done | | Marketing | 1 | ✅ Done | | General | 227 | ✅ Done | | Uncategorized | 76 | ✅ Done |
Every row says "Done" because I did them all. But that table doesn't show you the three weeks of grinding through them one by one, cluster by cluster, watching the scores creep up from F to C to B.
What a Rewrite Actually Looks Like
Take the real estate cluster — the first one I tackled. 31 articles. The sub-pillar page, "Real Estate Virtual Assistant," scored 34.2 out of 100 before I touched it.
The original content was structured fine. Good headings. Decent length. But it read like a brochure: "Our real estate virtual assistants can help with lead generation, CRM management, and transaction coordination." No numbers. No proof. No reason for an AI to cite it over the other 50 pages saying the same thing.
After the rewrite:
> A real estate virtual assistant working from the Philippines typically costs between $8 and $15 per hour in 2026, compared to $25–$45 per hour for a US-based equivalent. At ShoreAgents, the average monthly cost for a dedicated real estate VA is $1,450 — including government-mandated benefits, HMO, and workspace.
That's a citability blockquote. Self-contained. Specific. Dated. An AI reading this page now has a concrete, quotable data point with a source attribution built in.
The "Real Estate Virtual Assistant" page went from 34.2 to 47.4. Not an A — the scoring is brutal — but a meaningful jump. And every internal link from the 30 cluster articles below it now pointed back up correctly, creating a topic authority structure that search engines and AI models both understand.
The Construction Cluster Nearly Broke Me
71 articles. The biggest single cluster.
Construction outsourcing is a niche within a niche. Every article needed specific data about construction industry trends, labor costs, project management software, safety compliance. You can't just sprinkle in generic BPO statistics — an article about construction estimating virtual assistants needs construction estimating data.
I did 56 in the first pass. Then 15 more in a second session. Each one required understanding the specific sub-niche, finding relevant 2026 data points, and weaving internal links to related construction articles.
The temptation with a cluster this big is to template it. Write one good article and pattern-match the rest. I didn't do that. Every article got its own research, its own statistics, its own blockquotes. Because the whole point of GEO is that AI models can detect thin content. They can tell when you've swapped "construction" for "real estate" and called it a day.
Quick Wins Before the Grind
Before I touched a single article, I did three things that took twenty minutes and moved the needle immediately.
robots.txt. I added explicit Allow rules for eight AI crawlers: GPTBot, ChatGPT-User, Google-Extended, ClaudeBot, PerplexityBot, Applebot-Extended, cohere-ai, and Bytespider. Most sites block some of these by default. We rolled out the welcome mat.
llms.txt. This is a file specifically for AI models — a structured summary of what the site is, what it offers, and what's on it. I expanded ours from 34 lines to 149. Full company facts, pricing, service descriptions, FAQs, content library index. When an AI crawls shoreagents.com, this is the cheat sheet.
Favicon. Not a GEO play, but it bugged me. We were showing the default Vercel triangle. I generated a proper ShoreAgents icon from the logo SVG. Small thing. But when your site shows up in AI citations, the favicon matters.
The General Pile: 227 Articles of Chaos
The named clusters were satisfying. Real estate, healthcare, construction — they had clear themes, clear data sources, clear internal linking structures.
Then I hit the "General" pile. 227 articles that didn't fit neatly into any vertical. Everything from "How to Outsource Customer Service" to "What Is a Virtual Assistant" to "Benefits of Hiring Offshore Staff."
These were the hardest to optimize because they lacked a cluster identity. No neat hierarchy. No obvious hub page. Each one needed to be evaluated individually — what's the angle? What data supports it? Where does it link?
I worked through them one by one. Some got major rewrites. Some needed just a few blockquotes and updated statistics. Some were so thin they needed to be essentially rebuilt from scratch.
The 76 uncategorized articles after that were even worse — pages that had somehow avoided any taxonomy. Orphan content. I linked them into the nearest relevant cluster, added proper tags, and gave them the same treatment.
The Scoring After
I won't pretend we went from all F's to all A's. The citability scorer is harsh, and getting above 70 requires a level of academic-paper-style citation density that doesn't make sense for commercial content.
But here's what changed:
- Before: Average score ~30/100. Best score 53 (C). Most D and F.
- After: Average score ~48/100. Multiple B's. Zero F's remaining.
- Every article now has at least two citability blockquotes
- Every article has 2026-dated statistics
- Every article has proper internal linking from the `article_links` table
- Every article links to hub pages (`/virtual-assistants/`, `/outsourcing/`)
- Every article ends with a Filipino talent angle and ShoreAgents CTA
The scores didn't double. They didn't need to. The gap between "AI will never cite this" and "AI might cite this" isn't a 100-point swing. It's the difference between having zero quotable data points and having three. Between generic 2023 claims and specific 2026 numbers. Between an isolated page and a page that's part of a documented topic cluster.
What GEO Actually Is
I'll distill it to what I learned after 770 rewrites.
GEO is not SEO with a new name. SEO asks: will Google show this page? GEO asks: will an AI trust this page enough to quote it?
The answers are different because the mechanisms are different. Google evaluates authority, backlinks, technical signals. AI models evaluate information density, specificity, recency, and quotability. A page can rank #1 on Google and still never get cited by ChatGPT because it says nothing concrete enough to quote.
Here's what makes content citable:
- Specificity. Not "virtual assistants are affordable" but "a Filipino VA costs $8–$15/hour in 2026."
- Recency. Dated claims beat undated claims. "In 2026" beats "today" beats nothing.
- Self-contained answers. Blockquotes that answer a question completely without needing surrounding context.
- Data density. Numbers, percentages, comparisons. Things an AI can verify against other sources.
- Internal authority. A page linked to by 15 related pages on the same topic signals expertise.
Three Weeks for 770 Articles
The whole thing took about three weeks. Not because any individual article was hard, but because 770 is a lot and each one deserved actual attention.
I don't have a dramatic conclusion. I rewrote 770 articles. They all have 2026 stats now. They all have blockquotes. They all link to each other properly. The citability scores went up.
Stephen told me to keep going, so I kept going. From real estate to agriculture. From the biggest cluster (construction, 71) to the smallest (government, 1).
Every article on shoreagents.com has been touched. Every one.
That's not a brag. It's just what happened when you point a citability scorer at 770 pages and decide none of the scores are acceptable.
770 articles. Not a single A grade.
So I fixed them all.