SEO Forecasting Is Broken: One Question Fixes It
Queries an assistant can finish fell 64.2% from their peak. Everything else fell 4.5%. Same search index, same time period, same scoring rule. A 14.2-fold difference between two groups of keywords, and the reason conventional SEO forecasting has stopped working.
That gap is why most SEO forecasting is now producing numbers with very little relationship to next year, and this piece explains the fix. It takes an afternoon and needs no new software.
What every forecast quietly assumes
The standard model works like this. Take last year's traffic. Apply a growth rate. Adjust for seasonality. Add the content you plan to publish. Present.
Buried underneath is an assumption nobody states: that demand for a query persists, and the only variable is your share of it.
For twenty years that held. It has now stopped holding for one specific, identifiable group of queries. Any forecast that does not separate that group from the rest is averaging a collapse and a flat line into a number that describes neither.
The two groups, sorted before the data was pulled
I took six keywords and put them into two groups before looking at any trend data. That order matters. Sorting them afterwards would only prove I can draw a circle around points I had already seen.
Group A – substitutable. Questions an assistant can finish inside the tool you are already in. A syntax lookup. A formula. A method reference.
Group B – resistant. Questions where the answer is not the deliverable. Physical tasks, local intent, comparison shopping.
Then each keyword was scored against its own best calendar year, using monthly volume history.
Group A fell 64.2%. Group B fell 4.5%.
What that looks like keyword by keyword
The steepest fall was 74.7%, from a 2022 peak to 2026. No competitor was involved. Nobody out-ranked that page, built better links, or published a fresher version. The question stopped being asked.
At the other end, a local query sits 13.6% off its peak, which for something that seasonal is close to noise. Two of the Group B keywords had their best calendar year in 2026 – they are not declining at all.
Run a conventional forecast across all six as one group and you would have produced a modest single-digit decline, and been badly wrong about both halves in opposite directions.
Six keywords describe a mechanism, not a market. They show that substitutability predicts decline. They say nothing about how much of your business sits either side of the line. But the mechanism transfers, and you can test it this week.
The question that sorts your keywords
Take your top 200 queries by traffic. For each one ask exactly this:
Can a model complete this user's task without them leaving the tool they are in?
Yes means substitutable. Forecast a decline.
No means something has to be visited, compared, booked, measured or bought. That is where your traffic still is in 2028.
That is the whole method.
Two refinements while you sort:
Watch for good click rates on dying queries. One keyword in my basket has one of the best clicks-per-search figures I measured and is down 62.3% from peak. The casual majority left; the residue still clicks. Your click-through rate looks fine while sessions drain.
Score against each query's own peak year, not a fixed base. Fixing the base year flatters or punishes keywords depending on when they happened to be popular. Peak-year scoring treats every keyword by the same rule.
Why the industry-wide numbers cannot do this job
The macro figures are real and close to useless for planning.
In the US, 68.0% of Google searches now end without a click, up from 60.5% in 2024. The share that does click fell 19.1% in relative terms – roughly one click in five, in two years.
Apply that as a flat haircut and you will be wrong twice: too gloomy about the resistant half of your portfolio, and nowhere near gloomy enough about the substitutable half.
The macro number is an average of a 64% collapse and a 4.5% drift. Nobody's business is the average. The structural version of this argument, written for people running programmes rather than reading about them, is in the five shifts AI Overviews force on a search programme.
Is SEO dead? No, but that is the wrong question
The "is SEO dead" argument never resolves because both sides describe different halves of the same dataset.
In the same window that clicks per search fell 19.1%, Google's search revenue grew 17% year on year. Alphabet's own Q2 2026 CEO letter sets out the quarter in the company's words, and the picture there is of an engine in excellent health.
Both camps are quoting accurate numbers. Neither is describing your situation.
The useful question is not whether SEO is dead. It is: what share of my demand is substitutable, and what is my plan for it? That question has an answer, it is specific to you, and you can compute it.
Rebuilding your SEO forecasting model
Good SEO forecasting now needs two lines on the chart instead of one. Here is what replaces the growth-rate model.
Split the portfolio first. Substitutable and resistant, forecast separately, with different assumptions. One blended line hides the only thing worth showing.
On the substitutable line, forecast decline as the base case. Not flat, not modest growth. The observed rate here is a 64.2% fall from peak, and the mechanism driving it is not slowing: 74.1% of this year's model releases name reasoning, and the set of tasks an assistant can finish grows with each one.
On the resistant line, forecast normally. Growth, seasonality, content pipeline, share gains. The old model still works, because the old assumption still holds.
Report clicks per search next to sessions. Sessions can fall while your rankings improve, and that is very hard to explain to a CFO after the event. Clicks per search separates the market change from your performance.
Check what your tools can still see. Since late August 2026, Google routes result clicks through a redirect that blocks results-page scraping. Rankings are unchanged; rank-tracker visibility is not.
Prove causation where the money is. If a channel refuses to show up in last-click, forecast accuracy will not fix that on its own. Google's Meridian GeoX brings geo-incrementality testing to marketing mix models, which is the measurement answer to a channel you cannot attribute.
Re-run the split every two quarters. The line between the groups moves. A query that was resistant in 2024 – anything needing an image, a document or a screenshot – may not be resistant now. In 2023 this field shipped 25 text-only models; in 2026 it has shipped none.
Questions people ask about SEO forecasting
These are the questions searched most often alongside SEO forecasting, answered using the cohort data above.
How do you forecast SEO traffic accurately in 2026? Split your keywords into substitutable and resistant before you model anything, then forecast the two groups separately. A single blended growth rate now averages a 64.2% collapse with a 4.5% drift.
Why is my organic traffic falling when my rankings are stable? Most likely because the clicks left rather than the positions. Across the US, the share of searches producing a click fell 19.1% in two years. Check clicks per search alongside average position before assuming a ranking problem.
Is SEO dead? No. Google's search revenue grew 17% year on year in the same window clicks fell 19.1%. The engine is healthy; the traffic it exports is not. The question worth asking is what share of your own demand is substitutable.
Can you compare SEO forecasting tools? Any tool will apply a growth rate to historical data. None of them currently sorts your keywords by substitutability, which is the variable that now dominates. Do that split by hand first, then let the tool model each group.
How far ahead can I forecast reliably? Two quarters, then re-check the split. The boundary between substitutable and resistant moves every time model capability widens, and this field released a significant model every 2.98 days in 2026.
Cohort figures: Ahrefs monthly volume history for six US keywords, groups assigned before trend data was pulled, each keyword scored against its own best calendar year, captured 7 September 2026. Zero-click figures: SparkToro on Similarweb panel data. Release figures: a workbook of 261 model releases across 15 labs. Full method in the report, How AI Changed the Way People Search.
Tags
