I started TrafficWonk because I wanted to understand search visibility better.
That’s the simple version. The long version is I’ve been an SEO consultant for the best part of 20 years and I’ve been thinking about building this kind of SEO tracking system for about 12 years.
What has happened since is considerably less simple because the world has changed.
TrafficWonk has gradually moved from being an SEO visibility system towards something that models businesses, websites, people, capabilities, relationships, opportunities and changes in the real world.
Somewhere along the way I started asking a different question.
What would I build if the objective was not simply to build another useful SEO product?
What would I build if I genuinely wanted to create something that could change the world?
Not in the slightly ridiculous startup sense where changing the colour of a button apparently means you are ‘disrupting an industry’.
I mean something that could eventually have a positive impact on the lives of people across the world.
That question changed TrafficWonk.
It also changed the way I think about building with AI.
The biggest lesson so far is that AI makes building easier. It does not make knowing what to build easier.
In fact it can make that problem considerably worse.
What Is TrafficWonk Actually Becoming?
TrafficWonk is becoming an evidence-based system for understanding economic actors and what is happening around them.
That sounds considerably grander than where it started. Originally I was interested in visibility. Which businesses could be found? What did Google understand about them? Where were the gaps? Where were the opportunities?
Those questions are still important.
They are just no longer the whole problem.
A website is evidence about a business. Search results are evidence about a market. Companies House records can be evidence about an organisation. A person’s profile can be evidence about a role.
None of those things is reality itself.
That distinction sounds obvious once you say it.
It took me quite a lot of building to properly understand how important it is.
TrafficWonk increasingly sits between raw evidence and decisions.
The eventual system should understand entities, observations, relationships, capabilities, needs, constraints, opportunities, actions and outcomes.
More importantly it should understand how those things change.
That is where the World Engine starts to emerge.
Why Did A Search Visibility Tool Stop Being Enough?
The original TrafficWonk problem was essentially an SEO problem.
Businesses existed. Search demand existed. Google decided which businesses were visible. TrafficWonk could observe that environment and help explain it.
There is enormous value in that.
We’ve spent more than twenty years at GrowTraffic helping businesses become more visible. We’ve watched search evolve from ten blue links into local packs, featured snippets, shopping results, knowledge panels, AI Overviews and conversational answers.
Visibility matters.
But the deeper I got into TrafficWonk the more I realised visibility is only one manifestation of a much bigger problem.
Businesses constantly have needs.
They need customers. Staff. Suppliers. Software. Funding. Premises. Skills. Advice. Distribution. Manufacturing capacity. Marketing. Transport. Energy.
Other businesses and people constantly have capabilities which could satisfy those needs.
The difficulty is matching them at the right moment.
Search engines do part of that job.
Marketplaces do part of it.
Recruiters do part of it.
Salespeople do part of it.
Consultants do part of it.
A huge amount still depends on somebody knowing somebody.
Once I started looking at the problem through that lens TrafficWonk stopped looking like an SEO tool.
What Did The First Version Get Right?
The first versions got one fundamental thing right.
They observed.
That remains at the heart of everything.
Rather than asking an AI system to invent a picture of a company we could inspect websites, search results and other public evidence.
We could see what actually existed.
That matters because modern AI is remarkably persuasive.
Give an LLM an incomplete collection of information and it will often produce a beautifully coherent explanation of what it thinks is going on.
Unfortunately coherent is not the same as correct.
TrafficWonk forced me to think more carefully about the difference between evidence and inference.
That has influenced almost every architectural decision since.
The other thing the early system got right was scale.
Humans are brilliant at looking at one business.
We are considerably less brilliant at looking at hundreds of thousands of them every day.
Machines are good at that.
The interesting question becomes which pieces of evidence deserve expensive analysis and human attention.
That question turned out to be much more interesting than simply crawling another website.
What Did I Get Wrong?
Quite a lot.
Building software with AI has a strange psychological effect.
You can move so quickly that progress becomes addictive.
A new idea appears. You explain it to a coding model. Twenty minutes later there is a database table, an endpoint and something appearing in the interface.
It feels productive.
Sometimes it is.
Sometimes you have just industrialised a bad assumption.
I’ve built things which worked technically and were conceptually wrong.
I’ve produced enormous amounts of data which added almost no decision value.
I’ve had systems confidently classify businesses based on evidence that simply wasn’t strong enough.
I’ve seen automated processes produce exactly what they were programmed to produce whilst completely missing what I actually wanted.
That is an important distinction.
Software can be functioning perfectly and still be wrong.
The more capable AI becomes the easier it is to make that mistake at enormous speed.
Why Is ‘Unknown’ A Real Answer?
One of the most important lessons has been learning to respect ‘unknown’.
Databases do not particularly like uncertainty. Product teams do not particularly like it either.
We want yes or no.
Qualified or unqualified.
Good lead or bad lead.
Decision maker or not decision maker.
Reality is messier.
If I cannot prove that somebody is responsible for marketing at a business then the answer is not necessarily ‘no’.
It might simply be ‘we do not know yet’.
That sounds trivial. It isn’t.
Once an AI system starts converting missing evidence into negative evidence it begins silently corrupting the model.
The mistake then compounds.
Later decisions inherit confidence which was never deserved in the first place.
If I rebuilt TrafficWonk tomorrow I would make uncertainty a first-class citizen from day one.
Unknown is information.
What Changed When I Started Thinking In Evidence?
The system became considerably more useful when I stopped asking ‘what does TrafficWonk know?’ and started asking ‘what evidence does TrafficWonk have?’
Those are different questions.
A claim should have a source.
It should have a date.
It should have a confidence level.
Ideally it should have corroborating evidence.
It should also be possible for later evidence to contradict it.
That changes the architecture.
Instead of building a giant database of supposed facts you begin building something closer to an evidence network.
This matters far beyond TrafficWonk.
Anyone building AI systems which influence real decisions should be thinking about provenance.
Where did this information come from?
How old is it?
Was it observed directly?
Was it inferred?
What assumptions were used?
What would cause us to change our mind?
These questions are not glamorous.
They are considerably more important than another shiny AI feature.
Why Must Observation Be Separated From Interpretation?
One of the most useful distinctions I have made is between observation, interpretation and action.
Imagine TrafficWonk observes that a manufacturer ranks strongly for five products but has almost no visibility for a sixth product clearly listed on its website.
That is an observation.
‘They need SEO help for that product’ is an interpretation.
‘Caroline should contact the marketing director tomorrow’ is an action.
Those are three completely different things.
Early systems tend to compress them.
The machine sees something and immediately decides what it means.
That is dangerous.
The better approach is to preserve each stage separately.
An observation can support several hypotheses.
A hypothesis can be weakened by later evidence.
An action might require a much higher confidence threshold than simply recording the observation.
This creates another useful principle.
The evidence threshold should rise with the consequence of the decision.
Showing somebody an interesting pattern is relatively low risk.
Automatically contacting a managing director is higher risk.
Moving money would be higher again.
Good AI systems should understand that distinction.
Why Did The Sales Digest Become The First Real Test?
TrafficWonk currently has a wonderfully uncomplicated judge.
Caroline.
If the system gives our business development manager ten companies worth speaking to then it has created something useful.
If it gives her ten rubbish companies then all the clever engineering behind them is irrelevant.
That Daily Digest became far more important than I originally expected.
It forced TrafficWonk out of the comfortable world of technical metrics.
A model can have impressive precision scores.
A crawler can process millions of pages.
A database can contain enormous numbers of records.
None of that means Caroline can pick up the phone and have a useful conversation.
That is the test.
This is probably the single piece of advice I would give to somebody building an AI product.
Find your Caroline.
Find the person whose job becomes noticeably easier when your product works.
Then build for that reality.
What Has A Human In The Loop Taught Me?
Human feedback is often discussed as though humans exist to supervise AI.
I increasingly think that gets the relationship backwards.
The human is teaching you what the problem actually is.
Caroline can look at a company and spot things that are difficult to encode.
Timing.
Commercial plausibility.
Whether the evidence makes sense in context.
Whether there is actually something worth saying.
Whether an opening conversation feels credible rather than manufactured.
That feedback should not simply become another prompt telling the model to ‘think more like Caroline’.
That is a tempting shortcut.
It is also the wrong abstraction.
The better question is what underlying evidence Caroline is using.
Can we represent that evidence?
Can we identify the reasoning?
Can we see whether those signals generalise?
That is the difference between automating a person’s judgement and learning something reusable from it.
How Did The People Graph Change The Direction?
Companies do not make decisions.
People do.
That realisation pushed TrafficWonk towards what we now call the People Graph.
A company might need SEO.
That tells me almost nothing about who I should speak to.
Is there a marketing director?
Has somebody recently joined?
Is responsibility spread across several people?
Is the founder still making marketing decisions?
Does an external agency already appear to be involved?
Have senior people changed roles recently?
Suddenly the domain graph is not enough.
We need people, roles, relationships and time.
The obvious answer would be to become dependent on LinkedIn.
I do not want to build the World Engine on top of somebody else’s walled garden.
That has forced us to think harder about public evidence, identity resolution and corroboration.
That has been healthy.
Why Are People And Relationships Harder Than Domains?
A domain is comparatively easy.
growtraffic.co.uk is growtraffic.co.uk.
Humans are less cooperative.
Two people can have the same name.
One person can have several professional identities.
Job titles change.
Directors resign.
Businesses trade under brands which bear little resemblance to their legal entities.
Agencies operate websites they do not own.
Companies own domains they barely use.
Franchisees operate brands owned by somebody else.
This forced another important realisation.
Ownership, operation, management, access, verification and representation are different relationships.
A person having access to a website does not mean they own the business.
A company claiming a profile does not necessarily mean it owns the domain.
These distinctions sound annoyingly pedantic.
They are exactly the kind of distinctions machines need if they are going to act reliably in the real world.
Why Did State Change Become The Core Idea?
Static information tells you what something is.
State change tells you what might be happening.
That is considerably more powerful.
A business has 150 employees.
Interesting.
A business has just increased its workforce from 80 to 150 employees.
Much more interesting.
A company has a marketing director.
Fine.
A company has just hired its first marketing director.
Potentially significant.
A website ranks poorly.
Useful observation.
A website has lost half its search visibility over six weeks.
Now something has changed.
State change introduces time into the model.
It gives us a possible ‘why now?’
Salespeople instinctively understand this.
The difference between a cold database and a genuine opportunity is often timing.
That thinking pushed TrafficWonk beyond classification.
The goal became understanding changes in the state of economic actors.
Once I saw that I found it difficult to unsee it.
When Did TrafficWonk Become The World Engine?
There was not one dramatic moment.
It was more like gradually walking far enough from the original idea that I eventually turned around and realised I could no longer see it.
TrafficWonk had begun with websites.
Then businesses appeared.
Then markets.
Then people.
Then capabilities.
Then needs.
Then relationships.
Then observations.
Then changes.
Then opportunities.
Then actions.
Then outcomes.
At some point I started asking what happens if you keep following that logic.
What if the system could model enough of the economic world to understand who can do what?
What if it could understand who needs what?
What if it could detect when either of those things changes?
What if it could connect them?
That is the point where I started calling the bigger idea the World Engine.
What Do I Mean By A World Engine?
The World Engine is my long-term vision for a continuously updating model of economic capability, need, constraint, relationship and change.
It is not a giant crystal ball.
It is not an omniscient AI.
It is definitely not some bloke in Lancashire trying to create Skynet between SEO meetings.
The idea is much more practical.
The world contains enormous amounts of underused capability.
Factories have spare capacity.
People have skills which are not being fully used.
Businesses need services they do not know exist.
Suppliers could solve problems for companies which will never discover them.
People could collaborate but never meet.
Capital could be deployed better.
Knowledge could move faster.
The matching mechanisms are inefficient.
Search engines help.
Markets help.
Directories help.
Professional networks help.
But much of the world’s economy still operates through imperfect information.
The World Engine vision is about reducing that friction.
Why Start With Sales If The Ambition Is Global?
Because somebody has to pay for all this.
That answer is slightly less romantic than ‘changing humanity’.
It is also considerably more useful.
GrowTraffic has an immediate problem which TrafficWonk can solve.
We need to find companies we genuinely ought to be speaking to.
If TrafficWonk makes that process better then it creates revenue.
That revenue funds the next stage.
The rule I have gradually arrived at is simple.
Build one general primitive at a time and let GrowTraffic pay for it.
A better organisation model improves sales.
A better People Graph improves sales.
Better state-change detection improves sales.
Better capability modelling improves sales.
But each of those components could later be useful in completely different contexts.
That matters because I do not want to spend years building an elaborate theoretical machine which only becomes useful when the entire thing exists.
That is how grand visions become expensive hobbies.
Sales is the laboratory.
The World Engine is the destination.
What Would I Do Differently If I Started Again?
I would begin much closer to the decision.
I spent too much time thinking about what could be collected before thinking hard enough about what somebody needed to decide.
That is a common technology mistake.
We build the data lake first.
Then we stand beside our magnificent data lake wondering why nobody wants to drink from it.
I would start with the output.
What should Caroline know tomorrow morning that she does not know today?
What evidence would make that useful?
What minimum model is required to produce that evidence reliably?
Then I would work backwards.
I would also preserve raw observations much earlier.
Interpretations change.
Models improve.
Prompts change.
Business rules change.
Raw evidence is what allows you to reinterpret history.
If you throw it away too early you lose that possibility.
Why Would I Build Primitives Instead Of Features?
Features feel like progress.
Primitives create leverage.
A feature might be ‘find companies that need SEO’.
A primitive might be ‘represent an observed change in an economic actor with provenance, confidence and time’.
The first one solves one problem.
The second one can support thousands.
This is now one of the biggest tests I use when deciding whether something belongs inside TrafficWonk.
Are we building something which only makes sense for GrowTraffic’s sales team?
Or are we building a more general capability which happens to solve GrowTraffic’s sales problem first?
That distinction is crucial.
It prevents the system becoming an enormous collection of hard-coded assumptions about digital marketing.
It also stops the opposite problem.
You do not need to invent a complete universal ontology before you are allowed to build anything useful.
That road leads directly into what I call ontology swamp.
You can disappear into it for years.
Why Must Every Stage Earn The Right To Exist?
I am increasingly suspicious of technology roadmaps where all the value appears in year five.
A good system should begin creating value much earlier.
That value does not always have to be financial.
It could save time.
It could improve a decision.
It could reduce risk.
It could reveal something a human would have missed.
In TrafficWonk’s case I want it to make money for GrowTraffic because that creates a particularly useful discipline.
It means the system encounters reality.
Did somebody call the lead?
Did they reply?
Was there a meeting?
Was a proposal created?
Did we win the business?
Was TrafficWonk’s hypothesis right?
This creates an outcome loop.
Reality gets to vote.
That is incredibly important because AI systems can otherwise become self-referential.
The model predicts something.
Another model evaluates the prediction.
A third model writes a lovely report about how clever the first model was.
Everybody is delighted except the customer.
Where Does SEO Fit Into All Of This?
SEO is still one of the richest sources of observable commercial evidence on the internet.
Search reveals demand.
Websites reveal capability.
Content reveals expertise.
Links reveal relationships.
Structured data reveals explicit entities.
Search results reveal how Google interprets markets, brands and relevance.
That makes SEO data useful far beyond ranking reports.
The rise of LLMs makes this even more interesting.
ChatGPT, Gemini, Perplexity and Google AI Overviews increasingly mediate how people discover information.
Traditional rankings still matter.
But visibility is becoming more distributed.
A business might be discovered through Google.
It might be mentioned in an AI-generated answer.
Its content might become supporting evidence for an LLM.
Its brand may influence a recommendation without generating a conventional click.
That creates a new search landscape.
What Does AI Search Change About Visibility?
AI search makes machine comprehension increasingly important.
A page cannot merely contain keywords.
Machines need to understand the entities involved, the relationships between them and the claims being made.
That is where concepts such as entity salience, structured data, E-E-A-T and AI crawlability become more interesting.
Robots.txt rules matter.
Controls such as Google-Extended matter.
Clear authorship matters.
Consistent business information matters.
Original evidence matters.
Useful content still matters.
The irony is that the future of SEO may involve becoming considerably less obsessed with individual keyword rankings.
Zero-click search has already weakened the relationship between visibility and clicks.
LLMs push that further.
The question becomes whether your organisation is understood well enough to appear when a machine constructs an answer.
That is not completely different from what TrafficWonk is trying to do.
Both problems are fundamentally about entities, evidence, meaning and relationships.
What Could The World Engine Eventually Change?
The really ambitious version of this idea is not about helping GrowTraffic sell SEO.
It is about helping economic actors find each other more efficiently.
Imagine a manufacturer suddenly needs a specialist component.
Somewhere else a smaller manufacturer has exactly the capability required but is running below capacity.
Today those companies might never discover each other.
Imagine a growing company develops a skills gap.
A person living twenty miles away has precisely those skills but nobody knows the opportunity exists.
Imagine a council needs a particular capability.
A local supplier could provide it.
A technology company has created a solution.
A charity has discovered a problem.
A university has expertise.
An investor has capital.
A building is empty.
A machine has spare production capacity.
The world contains billions of these mismatches.
Most are invisible.
The World Engine vision is to make some of that invisible capability legible.
Then useful systems can help connect it.
If we get that right the impact goes far beyond marketing.
It becomes infrastructure for better decisions.
What Still Makes Me Sceptical?
Big visions are dangerous.
They make intelligent people start forgiving stupid decisions because everything can supposedly be justified by the glorious future.
I do not want that.
The World Engine should be attacked constantly.
Can the evidence support the inference?
Is the data legally and ethically usable?
Can the system explain why it believes something?
Does the model understand uncertainty?
Are we accidentally baking current assumptions into permanent architecture?
Are we building genuine intelligence or just an enormous collection of heuristics?
Does it produce anything useful?
Those questions need to stay uncomfortable.
I am a techno-optimist.
I genuinely think AI will allow us to build systems which would have been economically impossible even a few years ago.
But optimism without scepticism produces rubbish.
The trick is holding both positions at once.
This could be enormous.
This could also fail.
Both statements can be true.
So Where Does This Go Next?
The next step is not building the World Engine.
The next step is making tomorrow’s TrafficWonk better than today’s.
Better evidence.
Better people data.
Better understanding of relationships.
Better detection of change.
Better commercial opportunities.
Better feedback from what happened afterwards.
That is deliberately mundane.
I think the route to building something enormous is probably thousands of quite boring improvements which gradually become something extraordinary.
I still want to build something that changes the world.
I still want to build something which could eventually improve decisions for people and organisations almost everywhere.
The difference now is that I have a much clearer idea of how not to do it.
Do not start with omniscience.
Start with one useful decision.
Observe reality.
Preserve the evidence.
Admit what you do not know.
Learn from the outcome.
Then do it again.
That is TrafficWonk today.
The World Engine is what happens if we can keep doing it.
If you are building with AI inside your own business then that is probably the most useful place to start too. Forget the giant transformation programme for a moment. Find one real decision which happens repeatedly. Work out what evidence improves it. Build that loop. Make it genuinely useful. Then see where it leads.
If you’re interested in finding out more about TrafficWonk please head over to trafficwonk.com.
Frequently Asked Questions
Is TrafficWonk An SEO Tool Or An AI Sales Platform?
TrafficWonk currently sits across SEO intelligence, business data and sales intelligence. It uses observable evidence to identify organisations where there may be a credible commercial opportunity.
The longer-term architecture is broader than either SEO or sales. Those are simply the first environments where the underlying models can prove themselves against real outcomes.
What Is The Difference Between TrafficWonk And The World Engine?
TrafficWonk is the product being built now. The World Engine is the longer-term vision which has emerged from the concepts and infrastructure being developed inside TrafficWonk.
TrafficWonk needs to create practical value for GrowTraffic today. The World Engine explores what those same primitives could become if applied to economic actors, capabilities, needs and relationships at global scale.
Why Is Provenance Important When Building AI Business Intelligence?
Provenance tells a system where a piece of information came from, when it was observed and how much confidence should be placed in it. Without provenance an inferred claim can easily become indistinguishable from an observed fact.
That becomes increasingly dangerous when AI systems start recommending actions. The more consequential the decision becomes the more important it is to understand the evidence underneath it.
Why Should AI Systems Treat ‘Unknown’ Differently From ‘No’?
A lack of evidence does not prove that something is false. Treating missing information as a negative result introduces hidden assumptions into the model.
Good AI systems need explicit uncertainty. That allows future evidence to improve the answer rather than having to undo a false certainty which should never have existed.
How Can A Business Start Building Its Own AI Decision System?
Start with one recurring decision rather than trying to automate the entire company. Identify the evidence a good human uses to make that decision then capture the eventual outcome.
That creates a feedback loop grounded in reality. Once it works you can generalise the underlying components and gradually tackle more complex decisions.