Byword

Founder

Working as an independent consultant at a variety of brands, including ramp.com veloforte.com, causal.app, and many more.

Growth Consultant

Babylon Health

Biddable Lead

Brainlabs.

Account Director

Account Manager

Oxford Student Publications Limited

Director

Chairman

Goldman Sachs

Worked on projects:
a) Modelling settlement risk for $200 billion worth of daily FX trades carried out by the firm.
b) Predicting end of day asset hypes based on intra-day data.

Summer Analyst - Corporate Treasury Strats

CERN

Wrote code to display Fast Fourier Transforms (FFTs) of data taken from accelerometers placed on the Large Hadron Collider's (LHC) ATLAS Detector.

Technical Intern

(mack@byword.ai for Byword queries/support)

Something fundamental changed in AI citations last week.

We track citation patterns across AI engines at Trakkr (which sources get cited, how often, and how that shifts over time). 

For months, the picture was remarkably stable. Wikipedia held ~16% share. The same top 10 sources dominated.

Then around March 6, the whole distribution collapsed.

Wikipedia went from 16% of daily citations to 4%. 

In absolute terms - not just share - its citation count dropped 35%. Every major source we track (Forbes, TechRadar, LinkedIn, Reddit, Axios) showed the same cliff.

But here's what makes it interesting: total citation volume actually increased. It more than doubled. AI engines are citing more sources than ever - they're just spreading those citations across a dramatically longer tail.

Some numbers:

Daily citations went from ~25k to ~66k
Wikipedia: 4,231 citations/day down to 2,752
Wikipedia share: 16.3% down to 4.2%
Every top-10 source saw similar share declines

What this looks like to me: AI models are diversifying their sources. The era of a few dominant domains commanding most of AI's attention may be ending. The long tail is getting longer, fast.

If this holds, it has real implications for AI visibility strategy. Getting cited by AI is becoming less about being Wikipedia-adjacent and more about being the most relevant source for a specific query.

Could be a backend model update, could be a lasting shift. But it's the biggest structural change we've seen in 4 months of tracking this data.

Mack Grenfell

I rebuilt Byword from scratch over the last few months. 

The main reason was to move faster on the product, but taking the app off of bubble.io also gave me a lot more flexibility in how the marketing site, and its SEO pages, are built.

The problem I was trying to solve: 

- When you're going into crowded content spaces, everyone has the same AI models.

- The quality floor for written content is way higher than it was even a year ago. 

- So the question becomes less about the text and more about the page itself; what does it actually do for the person who lands on it?

Instead of generating articles, I built structured pages. The system works off a matrix: 34 content types crossed with 309 niches. But the scale isn't really the interesting part.

The interesting part is what those pages are. An SEO checklist page isn't an article about SEO; it's an interactive tool with checkboxes that save your progress, a progress bar, priority filtering. A content calendar page has a real timeline layout with days color-coded by activity type. A blog post ideas page has filtering by difficulty and category. 

These are essentially all micro-apps that happen to target search intent.

The key technical constraint is that AI never writes freeform. 

Every page starts with a strict JSON schema; the AI fills structured fields, purpose-built React components handle the rendering. Content and design are completely decoupled, so I can redesign every page without regenerating any content.

I use Gemini Flash rather than a frontier writing model. When you're filling schemas instead of asking for prose, you need something fast, cheap, and consistent; Flash handles structured JSON output natively. 

100 concurrent workers, the full corpus generates in under 3 hours. Titles are templated, not AI-generated, because a good format string is more reliable than a model at titles.

Results: 

Organic clicks went from around 1,000/week to around 6,000/week, with only half the pages indexed so far. Resource pages and interactive tools drive the bulk of it.

I think of this as full-stack SEO; not just writing content, but engineering the schema, the generation pipeline, and the frontend together as one system. Each content type gets its own renderer. Each page is a small application rather than text in a template.

Eventually this feeds back into the product. I want Byword to become the CMS where people build and ship these kinds of pages on their own sites; not just generating articles, but giving people the tools to engineer their SEO from the ground up. Instead of posting dev tickets to change their SEO fundamentals, you're talking to an agent that can work across your front and backend and put fixes in place in minutes.

Full write-up in the comments below.

I ran 920,000+ pairwise comparisons across ChatGPT, Claude, Gemini, Perplexity, Grok, Deepseek, Meta AI, and Google AI Overviews.

The question was simple: when you ask two different AI models "What's the best X?", how often do they actually agree?

The answer: less than half the time.

The #1 recommended brand matches across models on just 43.9% of queries. Claude and Deepseek are the closest pair at 36% agreement. 

Meta AI and Perplexity are basically off doing their own thing, agreeing with other models only 10-19% of the time.

The only queries where all 8 models unanimously agree? 

Massive category dominators. Nike for marathon shoes. Apple for smartphones. Amazon for online shopping.

That happens on 4.2% of queries.

For everything else (enterprise software, B2B tools, niche products) the brand a consumer gets recommended depends almost entirely on which AI they happen to open.

This is the largest public analysis of how AI models disagree on brand recommendations: full report drops Thursday.

I've been tracking AI crawlers for half a year now. 600,000+ visits, all logged through Trakkr.

Here's the visit distribution for every URL that got crawled:

1 visit: 88.5%
2 visits: 8.3%
3-5 visits: 2.4%
6-10 visits: 0.4%
0+: 0.3%

Even the outliers (P99) max out at 5 visits total.

This is so different from how we've learned to think about search. 

Googlebot revisits your pages weekly. If something's wrong with your title tag, or your content is thin, or your page speed is terrible - you fix it, Google recrawls, and your rankings update. 

The feedback loop is forgiving. You get to iterate.

AI crawlers don't work like that. They pass through once and they're gone. 

There's no second chance to make a better impression, no "let me update this and wait for a recrawl." Your page either communicates its value on that single visit, or as far as that AI model is concerned, it might as well not exist.

What's interesting is how this changes the calculus around content investment. In SEO, it's been perfectly rational to publish something mediocre, see how it performs, and improve it over time. 

A lot of successful content strategies are built on exactly that iteration loop. But if 88.5% of your pages get one shot with an AI crawler, that strategy doesn't transfer. The economics shift heavily toward getting it right the first time.

A lot of the AI SEO advice I'm seeing still assumes you can iterate your way into visibility. The data suggests that's not how this works.

Full data and study coming out later this week.

I've been tracking AI crawlers for half a year. 600,000+ visits across every site in the dataset, logged and categorised.

Here's the market share breakdown:

GPTBot (training): 57.2% - 329,572 visits
OAI-SearchBot: 15.1% - 87,155 visits
ClaudeBot: 3.8% - 22,074 visits
Meta: 3.2% - 18,234 visits
Amazonbot: 2.2% - 12,456 visits

OpenAI alone accounts for 72% of all AI crawl traffic. But market share is really just the starting point - the more interesting findings are in how these crawlers actually behave once they arrive.

88.5% of URLs get visited exactly once. 

Even at the 99th percentile, pages max out at 5 total visits. This is fundamentally different from Googlebot, which revisits weekly and gives you a continuous feedback loop. 

AI crawlers treat your pages as one-shot encounters. Either the content communicates its value on that first visit, or it doesn't get a second chance. 

The entire "publish, measure, iterate" playbook that's worked for a decade in SEO doesn't really apply here.

The entry point data is equally counterintuitive. 

21% of ChatGPT Search sessions start on blog pages. 1% start on the homepage. When you look at OpenAI's training crawler, 52% of all traffic lands at depth 3 - pages that are exactly 3 clicks from the homepage. 

AI isn't navigating your site hierarchy top-down. It's going directly to the content that answers specific questions, which in most site architectures sits several layers deep.

What surprised me most was how differently each crawler operates. 

ClaudeBot visits homepages 19.2% of the time vs GPTBot at 2.8% - completely different crawling philosophies. ChatGPT Search covers 76% of sites in the dataset at 1,362 visits per brand. ClaudeBot covers 56% at just 470 visits each - far more selective, almost curatorial. ByteSpider crawls more unique sites than any other bot despite ranking third in volume - wide reach, shallow depth.

Even the schedules diverge. OpenAI crawlers are 29-41% more active on weekends. ClaudeBot drops 8%. If you're doing server-side testing or rate limiting, the day of the week you run it on will affect which crawlers you're actually measuring.

The big picture takeaway for me is that we're not dealing with one "AI crawling" behaviour - we're dealing with several distinct systems that happen to share a category name. 

Optimising for "AI crawlers" as a monolith is roughly as useful as optimising for "search engines" without distinguishing between Google and Bing. The strategies are going to need to be more specific than that.

Full interactive study with methodology here: 
https://lnkd.in/eD6mkTJC

I've been analysing 600k+ ChatGPT Search logs to understand where and how it starts its crawl sessions:

Blog pages: 21%
Product pages: 3%
Homepage: 1%

Then I looked at the crawl depth data for OpenAI's training crawler:

Homepage (depth 0): 2.7%
1 click deep: 10.3%
2 clicks deep: 19.6%
3 clicks deep: 51.7%
4 clicks deep: 12.0%
5+ clicks deep: 3.7%

Over half the training crawl traffic lands at exactly 3 clicks from the homepage.

Both of these datasets are pointing at the same underlying pattern. 

AI isn't browsing your site top-down the way Google does - starting at the homepage, following your navigation, and indexing hierarchically. 

It's going straight for the pages that answer specific questions. Blog posts, help articles, deep product documentation. The stuff that tends to sit at depth 3 in most site architectures.

This makes sense when you think about what these crawlers are actually doing. 

They're building training data or fulfilling search queries. 

In both cases, they want information-dense pages, not navigational ones. Your homepage is a routing page. 

Your blog post about "how to choose the right running shoe for flat feet" is an answer to a question someone will ask an AI model.

The implication is a bit uncomfortable if you've been investing heavily in homepage experience and top-of-funnel brand pages. 

Those matter for humans navigating your site, but the data suggests AI crawlers are largely skipping them entirely and heading straight for your content layer.

If your best material is buried deeper than 3 clicks, it's probably still getting found. If it's shallow but thin, the depth won't save it.

Part of a larger study I'm publishing tomorrow - half a year of crawler data, 600,000+ visits.

I just published The Model Divergence Report - an analysis of how 8 AI models disagree on brand recommendations across 779K+ comparisons.

43.3% average agreement rate. Models agree less than half the time on which brand to recommend.

The most interesting findings:

- Meta AI and Perplexity are outliers - they agree with other models only 10-17% of the time
- Even the most correlated pair (Claude + Deepseek) only agrees 35%
- "Best-of" queries show the most divergence - exactly the queries where brands have the most opportunity
- Only 3.9% of queries get unanimous recommendations from all 8 models

Each AI model is effectively its own channel. Being visible on ChatGPT doesn't mean you're visible on Claude or Gemini.

Full report in the comments below

taplio

Mack Grenfell's Linkedin Analytics

Mack Grenfell

Performance ScoreMack Grenfell's LinkedIn Performance

Mack Grenfell's Best Posts (last 30 days)

Ready to grow your LinkedIn presence?

taplio

.css-6ytdur{background-image:linear-gradient(284deg, rgba(39, 43, 72, 1) 0%, rgba(33, 47, 183, 1) 100%);color:transparent;-webkit-background-clip:text;background-clip:text;}Mack Grenfell's Linkedin Analytics

Mack Grenfell

Performance Score.css-goub6s{font-size:var(--chakra-fontSizes-md);font-weight:500;color:var(--chakra-colors-gray-600);display:block;margin-top:var(--chakra-space-1);}Mack Grenfell's LinkedIn Performance

Mack Grenfell's Best Posts .css-1emoo84{font-size:var(--chakra-fontSizes-md);font-weight:400;color:var(--chakra-colors-gray-600);}(last 30 days)

Ready to grow your LinkedIn presence?

Mack Grenfell's Linkedin Analytics

Performance ScoreMack Grenfell's LinkedIn Performance

Mack Grenfell's Best Posts (last 30 days)