How this site measures things

Where the figures come from, for Australia

Every search-demand figure on this site was pulled from DataForSEO on 31 July 2026, for Australia specifically, in English. Each pull is stored as a file under research/ in the site’s repository, and each page names the file its figures came from. These figures ultimately sit behind the website design and development cost guidance published elsewhere on this site, for Australian businesses working out what a design and development quote should scope.

Four datasets are in use:

• research/national-volume-au.json — monthly Google search volume and keyword difficulty for 751 terms.
• research/cpc-competition-au.json — advertiser cost-per-click, competition band and competition index. This is what tells you whether anyone is paying to intercept a query, which is often a better signal of commercial value than volume.
• research/ai-vol-keywords.json and research/ai-vol-questions.json — monthly prompt volume to AI assistants, with twelve months of history for each term.
• research/serp-verification.json and research/serp-verticals.json — the actual search results for a term, classified by what kind of page ranks.

Nothing on this site is estimated from a tool’s projection or inferred from a competitor’s traffic report.

What “one entity” means

Search tools return a separate row for every phrasing of the same question. Treating each row as a page produces a duplicate-content farm where several of your own pages compete for the same query and all of them do worse than one would have.

Three patterns collapse most of that duplication, and all three were confirmed in the measured data rather than assumed:

“web design X” and “website design X” are the same thing. Of 92 measured pairs, 76 returned identical volume. Sometimes a pair returns identical volume, identical difficulty, identical cost-per-click and an identical competition index. When that happens, they are one entity that a tool has listed twice.

“X” and “X australia” are the same thing for a site that only addresses Australia.

“web design for plumbers” and “plumber web design” are the same thing — and the second phrasing is usually several times larger.

So one page per entity, with the other phrasings listed on it. Each page shows how many other phrasings resolve to it, and their combined volume is counted once, not summed. Where the phrasings differ in AI-assistant demand rather than Google volume, that is stated separately, because they are different measurements of different things.

What was rejected from the site’s scope

Two terms with real volume were dropped after the search results were actually looked at.

“Retail websites”: real volume, no commercial intent

This term turned out to return listicles and trade press — articles about examples of retail websites, and industry news — rather than anyone offering to build one. The volume was real; the intent was not commercial.

Solar: no meaningful Australian demand

This term returned no meaningful Australian demand under any of four query shapes.

Both are recorded because a methodology that only lists what it kept is a description of a conclusion, not a method.

The rejection reasoning was wrong, and this is the correction

Several other topics were dropped on the same evidence and should not have been. The error is worth stating precisely, because it is a reasoning failure rather than a data failure, and the data was fine.

When the search results for a term came back dominated by explanatory articles, tutorials and inspiration galleries rather than by suppliers, that was recorded as the topic failing a commercial-intent test, and the topic was dropped. Two separate questions were being collapsed into one.

The first question is what shape of page wins this search. A results page full of explanatory articles is strong evidence about that: it says a service page will not rank, and that anything published here has to be written as information.

The second question is whether the topic belongs on this site at all. The search results say nothing about that. It is answered by whether the topic is part of the subject the site covers.

Answering the first question and treating it as an answer to the second removed topics that are central to the subject: responsive design, web development, custom design, the mechanics of selling online. They were removed on the grounds that people researching them were not ready to buy. Being unready to buy is not a reason to be unable to find an answer, and a reference site that only covers the parts of its subject with buyers attached is not a reference site.

Those topics are being reinstated, written as information rather than as service pages. The measurement was never wrong; the inference drawn from it was.

The limits of these figures

Keyword difficulty is a model, not a measurement

This is set out at length below, because it is the single most useful thing on this page for anyone being sold search work.

AI-assistant volume moves quickly

The largest AI figure on this site fell by roughly a third between April and June 2026. A figure from that dataset is a snapshot of a system that is itself changing, and it is dated for that reason.

Volume figures are averages of a rounded, modelled monthly series

A term reported at 390 a month is not 390 people in any given month.

A null is not a zero

Several terms return no measurable volume at all, which means the tool could not measure demand, not that nobody searches. At least one page on this site publishes on a null and says so on the page.

Keyword difficulty contradicts itself, three times over

Every SEO tool publishes a keyword difficulty score, usually out of 100, and it is routinely presented to business owners as though it were a measurement. It is not. It is a model’s estimate, and on this project it has now disagreed with itself three separate times, on terms where the observable evidence says the opposite.

Show the numbers behind this chart

Keyword difficulty disagrees with itself across identical entities. Four phrasings of the Brisbane web design entity return identical Google Ads metrics — 2,400 searches a month, 35.33 advertiser cost per click, MEDIUM competition at index 38 — but keyword difficulty scores of 10, 23, 43 and 50. Two further pairs show the same pattern: tradie web design scores 8 against tradie websites at 40, and law firm web design scores 0 against law firm websites at 50.

Phrasing	Difficulty (modelled)	Searches/month	Advertiser cost per click	Competition (index)
website development brisbane	10	2,400	35.33	MEDIUM (38)
web development brisbane	23	2,400	35.33	MEDIUM (38)
website design brisbane	43	2,400	35.33	MEDIUM (38)
web design brisbane	50	2,400	35.33	MEDIUM (38)
tradie web design	8	390	25.02	LOW (16)
tradie websites	40	1,000	10.08	LOW (24)
law firm web design	0	320	15.06	LOW (5)
law firm websites	50	90	9.41	LOW (27)

Brisbane: one entity, four different difficulty scores

Four phrasings of one entity — “website development brisbane”, “web development brisbane”, “website design brisbane” and “web design brisbane” — return byte-identical Google Ads metrics: 2,400 searches a month each, 35.33 AUD CPC each, MEDIUM competition at index 38 each. Google’s own advertising system treats them as one thing and charges one price. The difficulty model returns 10, 23, 43 and 50.

Tradies: the smaller term scores easier

“tradie web design” returns difficulty 8. “tradie websites” returns difficulty 40. Both are Australian searches for a website for a trade business, and when the search results for both were pulled and read, both were dominated by agencies selling that service.

Law firms: the smaller term scores harder

“law firm web design” returns difficulty 0 at 320 searches a month. “law firm websites” returns difficulty 50 at 90 searches a month. The much smaller term scores as far harder, across largely the same ranking pages.

Those three cases have nothing in common except the metric that produced them. In each, the figures that come from somewhere observable — what advertisers actually pay, what competition band Google reports, which pages actually rank — are consistent, and the difficulty score is the only thing that moves.

The chart above plots all three. The top row is the Brisbane entity: four dots spread across half the scale, every one of them describing search demand that Google’s own advertising system prices identically.

Why this matters if you are buying search work. A difficulty score is derived, usually from the backlink profiles of pages currently ranking, weighted by a formula each tool keeps to itself. Two tools give different answers for the same term. The same tool gives different answers for phrasings of the same query. None of that is dishonest — a model is allowed to be a model. But it becomes a problem the moment a screenshot of a low difficulty number is used as evidence that a term is winnable, or a high one as evidence that it is not.

So on this site, difficulty is recorded and never relied on. Where it disagrees with the cost-per-click and the actual result page, the observable evidence wins. If a proposal you are reading rests its case on difficulty scores, the question worth asking is what the advertisers are paying and what is actually ranking.

What verification rejected

Measuring demand is the easy half. The half that decides whether a page gets written is reading what actually ranks, and on this project that has removed more demand than it has kept.

Show the numbers behind this chart

Demand declined after the search results were read. Eleven measured terms totalling 8,400 searches a month were declined after their search results were classified. The largest is wedding websites at 2,900 a month, rejected on a competition index of 91. Others were rejected because the results were government sites, listicles, design-inspiration galleries or career advice rather than businesses selling web design.

Declined term	Searches/month	Reason
wedding websites	2,900	competition index 91
ndis websites	1,900	ndis.gov.au ranks first
real estate websites	1,300	advertisers pay 1.66
small business website design	720	3 of 7 agency
retail websites	480	listicles and trade press
plumber websites	260	3 of 8 are round-ups
landing page design	260	0 of 9 agency
shopify website design	260	Shopify owns the page
website copywriting	260	2 of 8 agency
solar websites	40	no meaningful demand
web designer vs web developer	20	career-choice intent

Every bar is a term that was measured, wanted, and then declined. Some were rejected on advertiser signals before any results were pulled — wedding websites at a competition index of 91, real estate websites where advertisers pay under two dollars a click. Most were rejected after the results were classified and read: government portals, listicles, design-inspiration galleries, career advice, or a platform’s own properties occupying the entire page.

None of it was wasted. Knowing that a term is not a market is worth as much as knowing that it is, and considerably more than a page built on the assumption.

Volume is not value: cost per click and what an enquiry is worth

The industries on this site are not ordered by search volume, and the chart below is why.

Show the numbers behind this chart

Ranking by search volume and by advertiser value give different orders. Nine industry entities ranked twice: by monthly search volume, and by volume multiplied by advertiser cost per click. Construction moves from seventh by volume to fifth by value because advertisers pay 29.89 per click for it. Plumber moves from sixth to ninth because advertisers pay 7.79. Tradie web design is first on both.

Entity	Searches/month	Cost per click	Volume x cost per click	Rank change
tradie web design	390	25.02	9,758	unchanged
ndis web design	390	18.58	7,246	unchanged
dental web design	260	26.64	6,926	4 to 3
law firm web design	320	15.06	4,819	3 to 4
construction web design	110	29.89	3,288	7 to 5
not for profit web design	170	14.83	2,521	5 to 6
medical website design	90	24.79	2,231	9 to 7
manufacturing web design	90	22.30	2,007	unchanged
plumber web design	140	7.79	1,091	6 to 9

Multiplying monthly volume by what advertisers actually pay per click gives a different order. Construction moves up four places because a click there costs 29.89 while a plumbing click costs 7.79. That figure is not a price for anything this site sells — it is what a competitor is willing to pay to reach the same person, which is the closest available proxy for what one enquiry is worth.

It is a proxy, not a measurement, and it should not be read as one. But ordering by volume alone would put plumbing above construction and manufacturing, and on the available evidence that would be the wrong way round.

AI-assistant demand is falling

Show the numbers behind this chart

AI-assistant demand for commercial web design terms fell 46 per cent in ten months. Indexed to July 2025. Prompt volume across the measured commercial corpus fell 46 per cent by May 2026, while a definitional control question — what is a CMS — moved 0 per cent over the same period on the same measurement. June 2026 is shown but not relied on: it fell across every term including the control, which indicates an incomplete month rather than a real decline.

Month	Commercial corpus (prompts)	Indexed	Control: what is a CMS	Indexed
Jul 25	25,686	100	33,789	100
Aug	27,631	108	34,632	102
Sep	23,035	90	32,508	96
Oct	23,963	93	35,654	106
Nov	20,194	79	33,533	99
Dec	22,193	86	37,251	110
Jan 26	19,453	76	35,253	104
Feb	17,547	68	34,338	102
Mar	17,189	67	34,054	101
Apr	14,309	56	33,658	100
May	13,818	54	33,771	100
Jun 26 (incomplete)	10,091	39	27,462	81

Prompt volume across the commercial terms measured for this site fell 46 per cent between July 2025 and May 2026. That figure means nothing on its own — a fall could be a change in demand, or a change in how the measurement works.

The control is what makes it a finding. “What is a CMS”, a definitional question nobody’s business depends on, was measured the same way over the same twelve months and moved by a fraction of a per cent. A corpus that halves while a control stays flat is a change in the demand, not in the instrument.

Two cautions on reading it. June 2026 is plotted and not relied on: it fell 27 per cent across the corpus and 19 per cent on the control in the same month, which is what an incomplete period looks like rather than a real collapse. And this measures prompts, not citations. Somebody asking an assistant a question is not the same as an assistant citing a source, so prompt volume is a measure of interest rather than of opportunity.

A number a tool produced is not the thing the tool was summarising

The same lesson arrived a second time on this project, from a different direction, and it cost more.

Before any industry page was written, each candidate term had its search results pulled and classified, so that a page would only be built where the results were genuine agency pages rather than directories, job boards or round-up articles. That check killed “retail websites” — 480 searches a month, and a page one of listicles and trade press.

The classifier detected job boards, training providers and directories by their web address. Everything it did not recognise, it labelled “agency/other”. That name is the entire bug. The catch-all for unknown was reported as the category for the thing being proved, so every unrecognised result silently counted as evidence of commercial intent.

It surfaced on “ndis websites”, a term measuring 1,900 searches a month. The tool reported seven of seven agency results. Reading the actual titles: position one was ndis.gov.au, position four was the NDIS Quality and Safeguards Commission, position seven was a peak body, and position three was a listicle. Three of seven were agencies.

Re-running the corrected classifier across every pull on the project flagged two more. One was “retail websites”, already killed. The other was “plumber websites”, which had been adopted as a page’s canonical target on the strength of a reported eight-of-eight; re-classified, it returns three listicles out of eight, plus a trade association and an actual plumbing business. That page now rests on a different term, pulled and read for itself, and the decline is recorded on plumber web design.

Two habits came out of it, and both are now enforced by the build rather than by intention. A classifier must never assert the thing it is being used to prove — the catch-all is now called UNCLASSIFIED. And a page claiming its intent is verified must name the specific result-set file that was pulled for its own entity, which the build checks; prose saying “verified” is not evidence, and that is exactly where the error hid.

Nine pages on this site failed that second check when it was first run, because they cited a file covering other terms entirely. All nine now say plainly that their commercial intent is not verified. Their demand, difficulty and advertiser cost are measured; what actually ranks for them has not been looked at. That is a smaller claim than the one they were making, and it is the true one.

The failure that kept recurring

The honest account of building this site is not a list of the checks it runs. It is that the same mistake happened five times, in five different forms, and each time the thing that failed was not a check being absent — it was a check being asserted rather than executed.

The instances, because the pattern is only visible as a set:

A validation reported as clean that never ran

Seventeen files were reported as passing the content schema while missing a required field. The report was produced by summarising what should have happened rather than by running the validator.

A hundred edits measured against stale output

A build had failed hours earlier, leaving the previous version of the site in place. Every subsequent check read that old copy and reported on a version that no longer existed — passing work that had never been verified, and failing work that was already done.

A fuzzy match producing plausible wrong values

An attempt to reconcile page metadata against a plan used approximate title matching. It produced nine confident, wrong associations — an article about hosting latency mapped to an unrelated performance measurement. Every value looked reasonable in isolation.

A search that missed a leading slash

A check for orphaned pages used a pattern that silently failed to match two of them, reporting the site as fully connected when it was not.

Word counts estimated by eye

Two writers assessed the length of roughly ninety-five pages by feel. Measured afterwards, a quarter of the site sat below its own minimum, typically by two to three hundred words. Nobody was careless; nobody had run the counter.

A measurement mangled by the command line

A figure describing a quarter of the site was produced by a command typed directly into a terminal rather than saved as a file. The pattern it used to strip formatting contained backtick characters, which the shell interpreted as an instruction to run a command rather than as text. The pattern was silently altered before it ever ran, and the resulting count was wrong — reported, then corrected, in the course of writing about tools not being verified.

The general form is worth stating because it is not obvious: a command typed inline is not the same as a script saved to a file. Quoting, character escaping and substitution behave differently in the two contexts, and an expression that works correctly in a file can be quietly rewritten by the shell before execution. Nothing errors. The command runs and returns a plausible number. Anything whose result will be reported gets written to a file first.

What the instances have in common

In every case the check existed, and in every case it was cheap. What was missing was that it was cheap enough to run while working rather than afterwards. A validator that needs a full build gets run at the end, by which point the person who wrote the page has moved on and the correction falls to somebody who did not write it and cannot tell padding from substance.

The corrections were therefore not new rules. They were the same rules, made available a second after the work rather than an hour. A per-file checker reports frontmatter and word count in well under a second. A build refuses to grade output it did not just produce. And a schema rejects an over-long value instead of trimming it to fit.

Why this is on a public page

Because the alternative is a page listing guards and implying they were designed in advance. They were not. Most were written immediately after something got through, and several were written after something got through twice.

That is worth stating for two reasons. It is what actually happened, and this site is not much use if its own method page is the least evidenced thing on it. And it is the more useful account: anyone doing this work will hit the same class of problem, and “make the check runnable while the work is happening” is more transferable than any individual rule here.

The limitation is real and remains. None of this catches a page that is accurate, complete, well-sourced and simply not worth reading. That still needs a person.

What the machine catches, and what it does not

The checks described above are automated. They scan every finished page for a list of claims this site will not make — invented statistics, unverifiable social proof, superlatives, and any suggestion of client work in an industry where none has been delivered. A page that trips one of them does not get published; the build stops.

That works well for assertions. It does not work for implications, and the difference cost something worth recording.

A page in this section once carried the sentence: naming client work requires written consent, and consent that has not been obtained is not a detail to work around. Every automated rule passed it. There is no banned word in it. Read as a person reads it, it says that case-study-worthy work exists and is merely awaiting permission. The truth was that no such work existed, and permission had never been sought. “Not obtained” and “never sought” are very different statements. Only one of them was true.

A compliance reviewer found it. No scanner would have.

The same review found something more serious on the same page: a statement that audiologists are registered under the National Law and should display a registration number. They are not, and they have none — audiology is not yet in the national registration scheme, and a different regulator governs how those practices advertise. That was not an implication. It was a plain factual error in regulatory information, of the kind a practitioner could have acted on, and it survived every automated check because no rule in the list was about being right.

Two things came out of it. The pattern that could be mechanised was — consent language sitting next to case-study language is now blocked, along with several close relatives. And the general limit is now stated rather than assumed: automated checks constrain what a page asserts; they cannot tell whether it is true, and they cannot hear what it implies. Regulated-industry pages on this site go to a human reviewer before publication for that reason, and the pages that have been through that review say so in their own provenance.

There is a smaller version of the same lesson in the tooling. Both of the search-result classification failures described above, and both of the provenance failures, were cases of a number or a field being trusted in place of the thing it summarised. The correction in each case was the same: make the claim checkable against something, and where it cannot be, say so instead of implying otherwise.

The demand threshold, and what replaced it

The rule used to be that any page targeting an industry or a location had to measure at least 30 searches a month, with the search results inspected and confirmed commercial, or it could not publish.

That rule was guarding a real risk — doorway pages, meaning a page generated for an industry or a suburb from a template with the name swapped, saying nothing true of its subject. It was guarding it with the wrong instrument. What makes a page doorway content is not that few people search for it. It is that the page contains nothing specific to what it claims to be about. Search volume never measured that; it only correlated with it, and the correlation cost a large number of legitimate pages.

The threshold has been replaced by a three-fact rule. An industry or location page cannot publish unless it states at least three things true of its own subject and not of its siblings. Those things are: the regulator or licence that applies, the buying trigger, the proof asset that works there, the seasonal pattern, or the compliance limit. Fewer than three and the page does not exist. This is enforced by the build, not by review.

The reason to prefer it is that it tests the thing that actually matters. A suburb page that cannot name three things specific to that suburb is doorway content whether it measures ten searches a month or ten thousand — and under the old rule, the ten-thousand version would have sailed through.

Money pages are the exception and keep every original guard. A page that exists to route an enquiry still has to show measured demand and inspected commercial intent, because that page is asserting a market exists. An informational page makes no such claim.

Pages published below the old evidence floor carry a written justification in their source, rendered on the page under the evidence block, and it is there to be argued with.

Why there is no longer a commercial disclosure

Earlier versions of this site carried a disclosure at the top of every page that recommended anything, and a passage on this page explaining why.

The reason was specific. The site was operated by the same sole trader as a web design agency, enquiries were referred to that agency, and a site giving buying advice while routing leads to one of the suppliers has a self-preferencing problem under the Australian Consumer Law’s prohibition on misleading conduct. The disclosure existed because the conflict existed.

That arrangement has ended. This site is now a fully independent brand: no referral arrangement, no affiliate links, no sponsored placements, and no commercial interest in which supplier or platform a reader chooses. The exposure disappeared with the relationship, so the disclosure has been removed rather than left in place as decoration — a disclosure with nothing behind it trains readers to skip disclosures, which costs precisely when a real one appears.

The machinery that produced those disclosures is still installed and still runs on every build. It currently finds nothing to act on, and the check that guards it was rewritten so that it asserts the absence of undisclosed commercial links rather than the presence of disclosures on links that no longer exist. A check looking for something that cannot occur passes forever while testing nothing.

What has not changed: the author is a practising web developer with opinions formed by doing the work, and those opinions run through every page. Removing a financial interest removes one reason to distrust a recommendation. It does not make this site neutral, and no page here should be read as though it were. The current position is on commercial relationships.

Pricing, development cost and design cost figures are a separate question

The search-demand method above has nothing to do with the pricing bands. Those come from delivered projects and are governed by their own rules, set out in how the bands were assembled.