Field Notes
What Twenty Years of Algorithm Change Taught Me About What Survives
Every technique that works because of how a system is implemented has a short half-life. The proxy cycle, four things that never stopped working, and three things I got wrong.
Quick answer
Two decades of algorithm change have taught me one durable lesson: every technique that works because of how a system is implemented has a short half-life, and every technique that works because of what a system is trying to achieve survives. Directory submissions, exact-match domains, keyword density, article spinning, guest post networks, AMP — all of them worked, all of them died, and all of them died for the same reason. The people who kept their positions across twenty years were not the ones who adapted fastest to each change. They were the ones who never needed to.
I have been working in digital operations for over twenty years, and I have watched a specific pattern repeat often enough that I now treat it as a law rather than an observation. Each time, the surface details differ and the shape is identical.
This is not a nostalgia piece. It is an attempt to extract the pattern, because we are in the middle of another iteration of it right now, and most of the AEO advice being sold in 2026 is the 2004 playbook wearing new vocabulary.
The eras, and what each one killed
| Era | What worked | Why it stopped |
|---|---|---|
| Early 2000s | Meta keywords, keyword density, directory submissions | Trivially gameable. Signal became noise |
| Mid 2000s | Exact-match domains, anchor text at scale, link buying | The signal was hijacked faster than it was earned |
| 2011–2012 | Article spinning, content farms, link networks | Panda and Penguin. Quality assessment arrived |
| 2013–2015 | Semantic keyword coverage, entity association | Did not die. Became baseline |
| 2015–2019 | Mobile responsiveness, HTTPS, speed | Did not die. Became table stakes |
| 2019–2022 | Intent matching, topical authority, E-E-A-T signals | Did not die. Became the assessment itself |
| 2022–2024 | AMP, FAQ and HowTo rich results, content volume | Support withdrawn or devalued. Format arbitrage ended |
| 2024– | AEO tactics, prompt targeting, schema-for-AI | In progress. Most will not survive |
That distinction is the whole lesson and it took me an embarrassingly long time to formulate. Techniques do not die because they are old. They die because they exploited a proxy rather than serving the thing the proxy was measuring.
The proxy problem, stated properly
Every ranking system measures something it cannot observe directly. It wants to know whether a page is useful. It cannot measure usefulness, so it measures proxies: links, keywords, engagement, structure, freshness, entity association.
The cycle, every time:
1. System adopts proxy P because P correlates with quality Q
2. Practitioners discover P is measurable and manipulable
3. P gets optimised directly, decoupling from Q
4. Correlation between P and Q collapses
5. System devalues P, adopts P2
6. Everyone who built on P loses. Return to step 1.
Goodhart's law, running on a fifteen-year loop
with a marketing industry attached to it.
The people harmed at step 6 are always the people
who were most efficient at step 3.The last line is the part worth sitting with. Being excellent at exploiting the current proxy is not a neutral position — it is an accumulating liability that pays well until it does not. I have watched businesses with genuinely world-class execution at some specific tactic disappear inside eighteen months, not because they did it badly but because they did it extremely well.
My position
I was wrong about this for years, and specifically I was wrong in a way I now see repeated constantly. I believed that understanding the mechanism deeply enough meant you could stay ahead of it — that the skill was in reading the system faster than others. It is a seductive belief because it makes expertise the answer. What I eventually accepted is that mechanism knowledge tells you which techniques are fragile; it does not let you keep using fragile techniques safely. Understanding the proxy does not exempt you from the proxy collapsing.
Four things that have never stopped working
Across every era in that table, four things kept working, and they are unexciting enough that they get skipped in favour of tactics.
- Being the primary source of something. Original data, first-hand testing, a dataset nobody else holds. Every system that has ever existed rewards this, because every system is trying to find it. It cannot be arbitraged because producing it is the cost.
- Unambiguous identity. Being clearly and consistently identifiable as a specific organisation with a specific competence. This mattered as trust signals, then as entity resolution, and now as source selection. Same underlying property, three vocabularies.
- Technical legibility. Fast, crawlable, well-structured, semantically marked up. The specific requirements changed constantly; the principle of being easy for machines to process never has.
- Answering the actual question. Not the keyword, the question. Every improvement in query understanding over twenty years has moved systems closer to rewarding this, which means every improvement has advantaged people already doing it.
I am aware how bland that list reads. It is bland because it is the residue left after twenty years of removing everything that turned out to be temporary. The dull things are what remain.
Applying the test to what is being sold now
The useful thing about a pattern this consistent is that it is predictive. Ask of any current tactic: does this work because of how the system is implemented, or because of what it is trying to achieve?
| Current tactic | Implementation or objective? | My expectation |
|---|---|---|
| Writing self-contained, directly-answering sections | Objective — retrieval wants responsive passages | Survives. Will become baseline |
| Publishing original data nobody else has | Objective — every system seeks primary sources | Survives permanently |
| Consistent entity identity and corroboration | Objective — disambiguation is a real need | Survives permanently |
| Targeting specific prompt phrasings | Implementation — assumes the prompt is the query | Already partly wrong. Query rewriting defeats it |
| Schema markup as an AI citation tactic | Implementation — and a pipeline stage that discards it | Will not survive scrutiny |
| Volume publishing for AI coverage | Implementation — assumes retrieval rewards quantity | Will collapse, as every volume play has |
| Optimising toward a content score | Implementation — targets an average of page one | Produces convergence, not advantage |
Three things I got wrong, specifically
Since I am making predictions, I should show my record.
- I underestimated how long bad tactics keep working. Repeatedly. Techniques I judged obviously doomed persisted for years, and telling a client to stop something profitable because it was structurally unsound was correct on a five-year view and unhelpful on a twelve-month one. I have more sympathy now for people who take the short trade knowingly.
- I was too confident about mobile-first, too early. The direction was right and my timeline was badly wrong, which cost clients money in premature rebuilds. Being right about a trajectory is not the same as being right about a date, and I now separate those explicitly when advising.
- I dismissed structured data initially. It looked like a proxy that would be gamed and devalued like every other. It was not — because it feeds entity resolution, which is a genuine computational need rather than a ranking convenience. I had the right heuristic and applied it to the wrong category.
What is genuinely different this time
The uncomfortable part
I have said “this time is not different” for twenty years and been right every time. I am less sure now, for one specific reason: the previous eras all preserved the click. Every algorithm change redistributed traffic between sites; none of them removed the visit. Answer generation does. A system that satisfies the query without sending anyone anywhere is not a redistribution, it is a change in what the web is for. That is a different category of change and I do not think the historical pattern fully covers it.
Three ways the current shift breaks the pattern:
- The feedback loop is gone. For twenty years you could observe results, form a hypothesis, test it and see. Generated answers are non-deterministic, personalised and unlogged. The empirical method that made this field learnable barely functions.
- There is no guidance document. Google published guidelines, announced updates, and answered questions. Assistant vendors publish nothing about retrieval behaviour and have no incentive to start.
- The economics may not support the content. Search sent traffic that funded the pages search indexed. If citation replaces the visit, the corpus these systems depend on loses its funding mechanism, and nobody has proposed a workable replacement. I do not know how that resolves and I am suspicious of anyone who claims to.
What I would tell someone starting now
Learn the mechanism, not the tactics. Tactics have a half-life measured in months and mechanisms have one measured in decades — retrieval, ranking, indexing and disambiguation have been the same four problems the entire time, wearing different implementations.
Be suspicious of anything that works better the more you do it. Genuine quality has diminishing returns; exploits scale linearly right up until they invert. If a tactic rewards volume, it is a proxy exploit and it is on a timer.
And build the thing that would be valuable if there were no algorithm at all. That sounds like a platitude and it is the most operationally useful test I have. Original research, a genuinely useful tool, data nobody else has, an argument nobody else is making — those were valuable in 2005, they were valuable through every reassessment since, and they are the only assets I have seen survive a transition none of us predicted.
Everything else I have watched people build over two decades eventually stopped working, and the ones who lasted were not the fastest to adapt. They were the ones who had built something that did not require adapting.
How this plays out in the current transition is set out in core updates are not updates and ranking and citation are two different processes. The measurement problem it creates is in how to measure AI visibility without fooling yourself.
Frequently asked questions
Why do SEO tactics stop working?
Because they exploited a proxy rather than serving what the proxy measured. Ranking systems measure things they cannot observe directly, so they use proxies like links or keywords. Once practitioners optimise a proxy directly it decouples from the quality it represented, the correlation collapses, and the system devalues it. Goodhart’s law on a long loop.
Which SEO practices have survived twenty years of change?
Four. Being the primary source of something — original data, first-hand testing, a dataset nobody else holds. Unambiguous, consistent identity. Technical legibility, meaning fast and well-structured for machines. And answering the actual question rather than the keyword. All four are objective-aligned rather than implementation-dependent.
How can I tell if a current tactic will survive?
Ask whether it works because of how the system is implemented or because of what the system is trying to achieve. Implementation-dependent tactics have short half-lives. Objective-aligned ones survive and eventually become baseline. Self-contained answering passages and original data are objective-aligned; prompt-phrase targeting and schema-for-AI are implementation bets.
Does understanding the algorithm protect you from changes?
No, and this is a belief worth abandoning. Mechanism knowledge tells you which techniques are fragile; it does not let you keep using fragile techniques safely. Understanding a proxy does not exempt you from that proxy collapsing. The people most harmed when a signal is devalued are usually the ones who were most efficient at exploiting it.
Is the AI shift genuinely different from previous algorithm changes?
Possibly, for one specific reason. Every previous era redistributed traffic between sites while preserving the click. Answer generation can satisfy a query without sending anyone anywhere, which is a change in what the web is for rather than a redistribution within it. The historical pattern may not fully cover that.
What has been lost in the move to AI search?
The feedback loop. For two decades you could observe results, form a hypothesis, test it and see the outcome. Generated answers are non-deterministic, personalised and unlogged, so the empirical method that made search marketing learnable barely functions. There is also no equivalent of published search guidelines from assistant vendors.
What is the warning sign of a tactic on a timer?
It works better the more you do it. Genuine quality has diminishing returns — the tenth improvement to a page matters less than the first. Exploits scale linearly right until they invert. If a technique rewards volume in a way that quality does not, it is a proxy exploit rather than an objective-aligned practice.
What advice holds regardless of the algorithm?
Build what would be valuable if there were no algorithm at all. Original research, a genuinely useful tool, data nobody else holds, an argument nobody else is making. Those were valuable before any of the current systems existed, survived every reassessment since, and are the only assets that reliably outlast a transition nobody predicted.