Original research marketing: how to create evidence AI can cite

What counts as original evidence

Original evidence is something you found out first-hand that isn't already written somewhere else. Researchers call this primary research: you collect the data yourself, for the question you're asking. Secondary research reuses data someone else collected, such as an industry report or a government series. Both have a place, but only primary research gives a page something no other site can say.

Google's guide to optimising for generative AI search draws the line clearly. It calls a post like "7 Tips for First-Time Homebuyers" commodity content, based on common knowledge that could come from anyone, and tells site owners to create content from what they know and the in-depth experience they can bring. Rewording someone else's statistics into a fresh "report" doesn't get around that. Google's spam policies treat republishing other sites' content with only light changes, and without adding value, as scraping.

Google's helpful content guidance puts the test as a question: does the content provide original information, reporting, research or analysis? It also asks whether a page drawing on other sources provides substantial additional value instead of simply rewriting them.

The table sets out the research types a local business can realistically run. The whole process runs in eight steps, and the rest of this guide takes them in turn:

1. Choose one question customers already ask that your records can answer.
2. Write the method note before you pull any data.
3. Pull the records, strip out anything that identifies a customer, and fold small groups into wider ones.
4. Count, then have a second person check the counts against the raw records.
5. Write the main finding inside its limits, with the count beside every figure.
6. Ask the person who does the work to explain what the pattern means, under their own name.
7. Publish the study on its own page with the table, method, limits, author and date.
8. Rerun it on the same method when the next period of data arrives, and record what changed.

Pick a question your records can answer

Pick the question before you open the spreadsheet. A study question earns its place when it passes all five of these checks:

If a question fails the second check, fix the record-keeping first and come back once a full period has been logged properly. A study built on patchy records falls apart the first time someone asks how you got a number.

The question also has to be narrow. "What do customers in our area need?" can't be answered from any set of records. "Which fault did we find most often on hot water callouts last financial year?" can. The table gives starting questions by business type, with the record that answers each one and the trap to watch for.

1. Customers already ask it before they book.
2. Your records can answer it without guessing, because every job carries the fields you need.
3. The answer isn't obvious. If everyone in your trade would say the same thing, it's commodity content.
4. It stays inside one area and one time window, so the result means something specific.
5. You'd publish the result even if it looks bad for you. If you wouldn't, you're planning an advert.

Write the method and protect customers

A method note is a short, plain description of how you got each result. Write it first, then stick to it. Google's helpful content guidance points to product reviews that say how many products were tested, what the results were and how the tests were run, backed by evidence such as photographs. A trade study works the same way, and for a before and after test of search visibility itself, see how to design a local visibility experiment. Whichever research type you chose, the method note should cover:

Job records hold personal information, so strip it out before anything goes public. The OAIC's guidance on de-identification says information counts as de-identified when there's no reasonable likelihood of anyone being re-identified in the context where it's released. It warns that a rare characteristic, or other data that could be matched up with yours, can let someone be re-identified.

The same OAIC guidance lists techniques for lowering that risk. Two suit a small study: removing what it calls quasi-identifiers, such as significant dates or a profession, and what it calls rounding, which combines values into categories, the way an age can be shown as a range instead of a single year. In practice, report counts by suburb or region rather than by street, fold any group with only a handful of jobs into a wider one, and never publish names, addresses, phone numbers or photos that show a house number. If a survey will feed a public study, say so when you ask the questions.

Then treat each finding as a claim. The ACCC's advertising guidance for business says claims should be true, accurate and based on reasonable grounds, and that a business must be able to prove any claim it makes or advertises. A finding on your own website can read as a claim about your service. Keep the raw data and working files, so you can show how each figure was reached. Health and legal practices also have profession-specific advertising rules, so have whoever handles compliance check the study first.

• Question: the one thing the study set out to find
• Source: which records or survey, and who collected them
• Period and area: the dates covered and the suburbs or region
• Definitions: what counted in each category, such as what was logged as tree root intrusion
• Exclusions: any jobs or responses left out, and why
• Size: how many records sit behind each figure
• Changes: anything about how records were kept that changed during the period
• Review: the named person who checked the counts before they went live

State the limits, including awkward results

Every study has limits. Stating them plainly makes the findings easier to believe, and it stops a reader, or an AI summary, stretching a local result into a national rule. Most limits fall under these headings:

Keep the wording of every finding inside those limits. "Among the pre-1980 homes this business cleared drains for last financial year, tree roots were the most common cause" is a finding. "Tree roots cause most blocked drains" is a claim the data can't carry. Show the count beside any percentage, and if you set your figures against an official series, such as the ABS Counts of Australian Businesses, name the series and the release, and keep its definitions separate from yours.

Not every study gives a clean answer. If two causes are close, say so and report both counts. If a result runs against what you expected, or doesn't flatter your business, publish it with a short note on what might explain it and what you'll check next time. Label any explanation as a possibility and keep it out of the headline finding. A mixed result stated plainly is still original evidence, and it makes the rest of your findings easier to trust.

One low-cost study type belongs here: testing someone else's claim. Industry articles often repeat a statistic without saying where it came from or how many records sat behind it. Before you repeat one, find the original study, its sample size and its method. If your own records cover the same question, run the same count on them and publish what you found, with both methods described side by side. Where the results differ, the difference is your finding, as long as you name the likely reasons, such as a different region, period or customer mix.

• Selection: the data only covers people who called you, so it leans towards homes that already had a problem.
• Area: say which suburbs or region the jobs came from, and that other places may differ.
• Time: name the period, and note anything unusual about it, such as a very wet summer.
• Size: flag categories built on only a few jobs as early signs worth watching.
• Consistency: say who logged the data and whether categories were applied the same way throughout.

Build tables that people and AI tools can reuse

A clean table is the most reusable part of a study. Build it as an HTML table in the page text and treat any chart as an extra. Google's page on AI features and your website lists making sure important content is available in textual form among the practices that apply to AI Overviews and AI Mode, with images and video supporting the text rather than carrying it.

Google's Dataset structured data documentation says datasets such as tables or CSV files are easier to find in its Dataset Search tool when you describe them with structured data. A name and a description are required, and it recommends properties such as the creator, the licence, the period covered and the area covered. That's optional. Google's guide to generative AI search says structured data isn't required there, and there's no special schema.org markup to add.

A reusable table follows these rules, and the template below shows its usual columns:

Keep the raw records private, because they usually hold customer details, and publish only the grouped figures.

When you rerun the study, add the new period as a new column beside the old one. Keep the category definitions the same, and put a dated note under the table for any that changed. Put the period in the table title, such as "Blocked drain causes, Caloundra, 2025-26", so a copied table still says which year it describes.

• One idea per table, with a title that says what it shows, where and when
• Plain column headers that carry their units, such as "Jobs (count)" or "Days to repair"
• One row per category, in an order that means something, with a total row
• The sample size and period printed with the table, so it still makes sense when it's copied out
• No merged cells, and no meaning carried by colour alone
• A downloadable CSV of the same summary figures, if you're happy for others to reuse them, with the terms of reuse stated

Worked example: a drain study from question to rerun

Say a plumbing business in Caloundra wants to answer a question its office hears every week: why do drains keep blocking in older homes? The business is invented for this example, and so is every number below. The figures show the shape of a finished study, not a benchmark for any real area.

Step one is the question: among blocked-drain jobs in Caloundra and nearby suburbs last financial year, what caused the blockage, and did the pattern differ by the age of the house? It passes the five checks. Customers ask it, every job sheet records the cause and the build era, and the owner agrees in advance to publish the result whatever it shows.

Step two is the method note, written before anyone opens the job system. It defines each cause, so tree root intrusion means roots seen inside the pipe on a camera inspection. It excludes jobs without a camera inspection, sets the build-era bands at before 1980 and 1980 or later, and names the office manager as the person who checks the counts.

Step three is the pull. Names, street addresses and photos stay in the job system, and the study works only with cause, build era, suburb and month. Two suburbs with only a few jobs each are folded into a wider "other nearby suburbs" group.

Step four is the count, shown in the table with illustrative numbers. The headline written from it reads: "Among 140 camera-inspected blocked-drain jobs this business attended in Caloundra and nearby suburbs in the 2025-26 financial year, tree roots were the most common cause in homes built before 1980, while in newer homes grease and fat build-up (24 jobs) and wipes (21 jobs) were close." That sentence carries the count, the area, the period and the split, so it still holds when it's quoted on its own.

Step five is the limits paragraph. The data only covers people who called this business, jobs without a camera inspection were left out, and the build era was recorded by the plumber on site. The collapsed or misaligned pipe row rests on 14 jobs, so it's reported as a count with no comparison drawn, and grease and wipes in newer homes are only three jobs apart, so the page reports both counts instead of naming a winner. The "other or unclear" row stays in, because hiding it would make the categories look tidier than they were.

Step six adds a named interpretation. The licensed plumber who ran most of the camera inspections explains, under their own name, what they think the root pattern says about older pipework. The table stays the evidence, and anything the plumber says beyond it is marked as their view.

Steps seven and eight are publishing and the rerun. The study gets its own page with the headline first, then the table, method note, limits, the plumber's commentary, an author and a published date. A year later the business runs the same count, adds a 2026-27 column beside the old one, and notes under the table that it began logging roots found at the boundary trap as a separate category partway through the year, so readers can see why that row moved.

Publish, share and check whether it worked

Give each study its own page. Put the main finding in the first two sentences, then the table, then the method note and limits, with a named author and a clear published date. The guide to AI citations and supporting evidence covers sourcing and corrections. Larger studies are often split into layers, and a local study can do the same on a smaller scale:

Check that search tools can reach the page. Google says a page must be indexed and eligible to show with a snippet before it can appear as a supporting link in AI Overviews or AI Mode. Google also requires the site to be included in Search generative AI features. That's the default, but check in Search Console, under Settings, that it hasn't been switched to exclude. Being reachable is only the first step, as why a crawl does not prove an AI citation explains.

Then tell the people who might use the finding. A local newspaper, an industry association, a supplier or a council program working on the same problem may quote it or link to it. Send the headline finding, the table and a link to the study page. Once a finding is quoted elsewhere you can't control the wording, so the page it links back to needs the method and limits in full.

Check what happened at set intervals, such as monthly, using the signals in the table. No one can guarantee an AI product will cite your research. What you control is whether the evidence is real and whether a careful reader can check it, which is the groundwork every AI SEO plan depends on. For reruns, the process for keeping AI visibility content current sets out how to update the page and keep a revision history.

• The main study page: headline finding, summary table, limits and commentary
• A method page, if the method note is long or shared by several studies, linked from each study that uses it
• Suburb or service slices, published only where each slice has enough jobs to stand on its own
• A short FAQ on the study page answering the questions the finding raises
• A CSV of the grouped figures, with the reuse terms stated