How to design a local SEO experiment you can trust

Why before-and-after numbers can fool you, and which design to use

A typical SEO check compares this month with last month, but plenty else moves in between. Google's guide to debugging traffic drops lists algorithm updates, technical problems, seasonal demand and site moves as causes of change, and says small position changes can happen at any time without you doing anything.

A control group fixes that: a set of pages or Business Profiles like your test group that don't get the change. If both groups move together, the market moved. If only the test group moves, your change is the likely cause.

Google Ads can split searchers at random: its cookie-based experiments assign each user to the original or the trial campaign. Organic search can't, because each page has one indexed version to rank. Google's testing guidance has alternate test URLs point back to the original, and notes Googlebot generally sees only the version served without cookies. So in SEO you split pages or places, not people.

Not every local business has twenty similar pages to split. The table compares the designs open to you, from strongest to weakest. Pick the strongest one your site can support, and name it in the plan, because the design limits what the result can prove.

Step 1: choose what to test and write the plan

Start with changes that are cheap to make, easy to undo and repeated across many similar pages. They give the cleanest test and the biggest payoff if they work. Search Console's Performance report points to good candidates: pages with plenty of impressions but few clicks, where the listing is seen but not chosen, and suburb pages built on one template, where a change can be split across near-identical pages.

Then write a hypothesis, not just a change. 'Adding a short local FAQ block to suburb pages will lift impressions for question searches, because the pages answer none of them now' says which measure should move first and in which direction. Without that, any movement in any number can be called a win after the fact.

The table pairs common local changes with the first measure to watch. Test one change at a time: rewrite titles and add a local FAQ block in the same week and you'll never know which did the work.

1. The one change, described so someone else could repeat it
2. The hypothesis: why the change should work, which measure should move first and in which direction
3. The design, and the test and control pages or profiles, listed by URL or profile name
4. The main measure and the supporting measures
5. The baseline dates, the test start date and a fixed end date
6. What counts as a win, a loss or no clear change, written as numbers before the test starts
7. What you'll do in each case: roll out, roll back or retest
8. Who can sign off other edits to those pages while the test runs

Step 2: build a baseline and set up the measures

A baseline records how both groups behaved before the change. Make it at least as long as the test, and check the same weeks last year for seasonal swings: Search Console's Performance report can show the last 16 months. Google Trends shows relative interest on a 0 to 100 scale, so use it for the shape of the year and ignore it for volume. The local SEO seasonality guide covers those patterns.

Know the order in which measures move. A change has to be crawled before it can be indexed, indexed before it can earn impressions, and shown before it can earn clicks, calls or forms. So the early measures are crawl dates and impressions, then average position and clicks, and the leads come last. A test where impressions rose but clicks haven't yet isn't a failed test; it may simply be early. Name one layer as the main measure in the plan and read the others as supporting evidence.

Map pack rankings need their own method. Google's local ranking help lists relevance, distance and prominence as the main factors, and says that when a searcher doesn't share their location, Google uses what it knows about it. A single check from your office therefore shows one view out of many. Choose fixed check points across the service area, such as the centre of each test and control suburb, and check the same search terms from those points on the same day each week, on the same device type and signed out. A geo-grid rank tool does the same job on a schedule. Record the position at each point rather than one average, because a change can lift rankings near the business and do nothing at the edge of the service area.

Then set up the remaining measures before the baseline starts.

• In Search Console, filter Pages with a custom regex matching only the test URLs, then repeat for the control URLs. Google warns that filtering by query or URL can affect report totals because of data truncation and the omission of anonymised queries. So use the same filter method for both groups, and never compare either with the unfiltered property total.
• View the data weekly: Google says weekly or monthly views neutralise the day-of-the-week effect of comparing a Monday with a Saturday.
• Mark calls and forms as key events in Google Analytics before the baseline begins. Google says a key event affects reports from the time it's marked and doesn't change historic data, so tracking added mid-test leaves no baseline. The call tracking and conversion reconciliation guide shows how to check those numbers.
• For Business Profile tests, Google's performance data includes views, searches and actions from organic results and Google Ads, and counts a person only once a day. Note the start-of-month update of the searches figure when you set the baseline length.
• Keep a dated spreadsheet of every weekly reading for both groups. Report screens change and data ages out; your own record doesn't.

Step 3: choose test and control groups that match

The two groups need to behave alike before the change. Pair similar pages, then flip a coin within each pair to decide which gets the change. Hand-picking tends to put your favourite pages in the test group, which tilts the result.

Take extra care with distance. Google's Business Profile help lists it as a local ranking factor, so a test group of suburbs near the business and a control group of distant ones will drift apart whatever you change.

Before you start, chart both groups week by week across the baseline. If the lines already drift apart, re-pair them. If you only have a handful of pages, don't force a split that leaves two pages on each side: use a staggered rollout or the before-and-after design from the first table, and treat the answer as a first read.

Step 4: time the test and log every confounding change

Give the change long enough to register, and keep Google's ranking updates out of the window where you can. Google's SEO Starter Guide says some changes might take effect in a few hours and others could take several months, and suggests waiting a few weeks before judging a change.

A confounder is anything besides your change that could move the numbers. You can't stop them all, but you can log them and hold the ones you control steady. Search Console lets you annotate Performance charts with notes of up to 120 characters on a date, up to 200 per property. Google deletes annotations older than 500 days, so keep your own change log too.

• Start the clock once Google has seen the change. URL Inspection shows when Google last crawled each page, and Google says indexing after a request typically takes a day or so but can take much longer.
• Check Google's Search Status Dashboard before you begin and weekly after. It lists ranking updates and their dates; past rollouts have run from under a day to several weeks.
• If a core update lands mid-test, Google recommends waiting at least a full week after it finishes before analysing, then comparing with a week before it started.
• Don't start or finish on the edge of your busy season, when demand shifts fastest.
• Set the end date in advance and stick to it. Stopping on a good first week locks in a high point that may fade.

Step 5: read the result, with a worked example

Judge the result by the change in the gap between the groups, and only call it a result if that change is bigger than the normal week-to-week wobble in the baseline. The example below uses illustrative numbers, not benchmarks.

Say a plumbing business in Caloundra has 24 suburb pages on one template. It pairs them by baseline clicks and distance from the workshop, flips a coin in each pair, and adds a local FAQ block to the 12 test pages. It records eight baseline weeks, makes the change, confirms in URL Inspection that Google has recrawled all 12, then records eight test weeks.

Both groups rose, so part of the lift is the season: the control group's 7 extra clicks a week would have arrived anyway. The change is worth about 10 clicks a week across the 12 pages. The test-period gap of 12 sits above the baseline's widest week of 7, and in this example the weekly gap stayed above the baseline's highest gap of +7 in six of the eight test weeks, so ordinary noise is an unlikely explanation. Had the test gap come in at 6, inside the baseline range, the honest verdict would be no clear change.

Now the leads. Say the test pages produced 11 tracked calls and forms across the test weeks against 8 for the control pages. Counts that small swing on one busy week, so treat them as supporting the click result rather than proving it, and keep counting after rollout.

AI answers can be read the same way. Search Console's Generative AI performance report shows impressions when links to your site appear in AI Overviews or AI Mode. Google's help says it rolled out to all websites worldwide as of 31 August 2026, yet lists three reasons a site may not see it: the rollout is still reaching some properties, the site hasn't had enough impressions in generative AI features, or the site is excluded from Search generative AI features. Group it by page and total the test and control URLs separately. The Google AI impressions guide explains what those numbers mean.

For ChatGPT and other assistants, run a fixed prompt list on a set schedule and log every run, including ones that name no business. OpenAI's help says ChatGPT may use an approximate location from your IP address, and saved memories when rewriting a search with memory on, so keep the device, location, account and memory settings the same. Treat each run as one sample: run every prompt several times and compare how often you're named before and after. Comparing AI visibility reports covers the settings to hold steady.

No one can guarantee a change will lift rankings or AI mentions. A well-run experiment tells you whether it did on your site, this time.

Roll out, roll back, write it up, and the tests never to run

Some so-called experiments break Google's rules or the law, whatever the numbers show. Google's Maps user-contributed content policy says contributions should reflect a genuine experience at a place or business, and that fake engagement is not allowed and will be removed. It bans offering incentives such as payment, discounts or free services in exchange for reviews, and warns that harmful behaviour can lead to action ranging from suspending account privileges to account termination. Services that sell fake searches, clicks or visits to a profile are selling engagement that doesn't reflect a genuine experience, and Google's local ranking help says there's no way to request or pay for a better local ranking.

In Australia, the ACCC says it's against the law for a business to create fake or misleading reviews or to arrange for others to do so, and that offering incentives for positive reviews risks misleading consumers and breaching the law. The Google review strategy guide covers asking for reviews the right way.

Business Profile tests have one more trap. Google's guidelines for representing your business say including unnecessary information in the business name isn't permitted and could result in suspension, so adding suburbs or service words to the name is not a test worth running. The Business Profile suspension guide covers what happens if a profile is suspended. On the website, the rule from Google's testing guidance holds: never show Googlebot something different from what people see.

• Win: roll the change out to the control group, then watch whether its numbers follow the same pattern. That second change works as a free retest.
• Loss: revert the test pages, confirm in URL Inspection that Google has recrawled them, and log the date.
• No clear change: keep the change on the test pages only if it helps visitors anyway, and don't roll it out on the strength of the test. Retest with more pages or a longer window if the idea still matters.
• Write each test up on one page: the hypothesis, the design and groups, the dates, the confounders logged, the numbers for both groups and the decision.
• Keep the write-ups together, including the no-change results. They stop you rolling a dud across every page and give the next test a starting point.