Average position in an AI report: check the Search Console scope first
Review AI analysis of Search Console: average position, query mix, filters, canonical URLs and limits on retrieved rows.
Article language: English

In this article
Average position fell from ninth to almost fourteenth. An AI report calls it a loss of visibility and recommends urgent page changes. Yet neither of the two observed queries changed position. The query for which the page had always ranked lower received more impressions. One average can tell a persuasive story about a ranking shift that never happened within either group.
This is a fictional example. Before interpreting a Search Console report, establish which observations make up the number. AI can help describe the material, but missing filters do not justify a story about a search algorithm. Correct arithmetic without context can still send a team to work on the wrong problem.
What does position mean?
Google explains that average position in Search Console concerns the highest result for a page or property, depending on aggregation, averaged across impressions. It approximates result placement; it is not a fixed position seen by every searcher or a substitute for one browser check.
In a working dataset, the first query has 100 impressions at position 3 and the second 100 at position 15. The overall average is 9. In the next period the first still has 100 impressions at position 3, while the second has 900 at position 15. The overall average is now 13.8. The shares changed, although the position within each group did not.
Before saying a page fell, distinguish changes within comparable groups from changes in their shares. The larger share of the second query needs its own business interpretation. It might concern an area on which the company only recently began publishing. A worse average alone does not show that the new visibility is harmful, nor that everything is working well. It identifies a question to investigate.
Prepare two views: the result for the selected scope as a whole and a comparison of particular queries, pages or groups relevant to the decision. Label them differently. Do not call a small selection of phrases the average for the entire site, or present all queries as the result of one service. Readers should be able to see what lies under each heading.
Keep groups present in only one period in the comparison. If you use only common rows, say so. New and disappearing queries may matter; combining everything without distinction can conceal a change in composition. Both contexts are needed, rather than whichever gives a prettier arrow in a presentation.
How can the report's scope be reconstructed?
Record the Search Console property, periods, search type, filters and grouping, especially country, device, query and page. Mobile traffic in one country is not comparable with all results merely because both spreadsheets have a position column. Remove scope differences or describe them as part of a different question.
The Google performance report guide distinguishes chart and table metrics and grouping by dimension; search type and date range can also be changed. An intake record should therefore preserve the configuration as well as the displayed numbers.
Start with a question such as: “How did impressions and clicks for these service pages change for queries related to that service, within comparable scopes?” It is easier to assess filters against that question. A vague “assess SEO” encourages a mixture of the whole domain, one article and an arbitrary group of terms, producing recommendations without answering the business need.
If you separate branded queries, preserve the definition: name variants, matching method and the date a rule changed. Do not compare today's split with an earlier one that assigned misspellings or a product name elsewhere. The difference can arise during preparation even when source data has not changed; the model should not attribute it to new search-engine behaviour.
For short periods, describe the calendar and completeness assumed for the assessment. Do not automatically explain every deviation as seasonality. Without evidence, keep it as a hypothesis. A report can list possible causes for investigation, but must separate them from observations in the dataset.
Did the model receive every row it discusses?
The Search Analytics API returns data for specified dimensions and filters within retrieval limits. Google says it does not guarantee every row, only top rows. Fetching subsequent pages therefore does not justify claiming that the analysis covers every user query.
Ask the integration provider to retain parameters, the number of rows received and the way retrieval ended. Distinguish all responses obtainable by that method from complete knowledge of search. An error on the second part must not silently turn the report into an analysis of the first. If an earlier successful dataset is retained, show its date rather than calling it fresh.
Preserve the meaning of an empty result. A missing row does not by itself mean position zero or that nobody entered the query. “No observation in the material received” may be the right answer. The model should avoid filling that gap by guessing from industry knowledge.

Do not let AI average existing row averages indiscriminately. First determine whether groups can be combined and the right basis for aggregation. Different impression counts change the combined figure in the opening example; adding two positions and dividing by two would answer a different question. Record whether the total came directly from the report or was calculated from a defined dataset.
For acceptance, also choose a case where the integration receives less data than expected. On synthetic input, decide in advance whether it should stop or present a limited result with a clear warning. Check the final summary, not only retrieval logs: a log error and a presentation claiming completeness contradict each other. This is a proposed test of your own system, not a claim that a product was tested here.
The earlier guide to percentages and units in an AI-written A/B report (in Polish) (in Polish) helps check calculations. This article does not assess an experiment. The shared task is reconstructing what was counted. A calculable percentage change does not turn an observation into a causal conclusion.
Which URL received the result?
Google explains attribution of performance data to the canonical URL. Usually this is the address selected by Google to represent page variants, with documented exceptions. Do not automatically expect separate figures for every technical URL variant or sum them without checking the attribution method.
When a model reports no result for a URL, compare that address with the one in the report. Record the domain variant, path and known redirects. When an old article has a new address, map the content rather than matching only similar titles. Preserve the transformation instead of silently replacing old addresses in the source file, so the page actually assessed remains traceable.
Performance analysis is not a technical acceptance test. Google's Search Console getting-started guide describes separate uses for performance, indexing information and URL inspection. They answer different questions. A number's presence or absence in one dataset does not replace checking a particular URL's current status.
The technical SEO acceptance guide (in Polish) (in Polish) addresses publication and consistent URLs. A page can be accessible yet have too few observations for a meaningful comparison. Do not promise indexing or rankings because a report has been repaired. Better description restores a basis for decisions; it does not automatically change audience or search-engine behaviour.
What should the sentence approved by the company say?
A useful summary combines observation, scope and limits: “The overall average worsened as the second query's share grew; the positions of both observed groups stayed the same.” The reader need not accept a diagnosis of an algorithm. Next, check whether the query gaining impressions corresponds to the offer and brings useful visits.
Do not equate a click with a received sales enquiry or a sale. That later stage has its own sources and limitations. The guide to B2B sales attribution and multiple contacts (in Polish) (in Polish) explains why a single contact record cannot describe an entire purchase journey. Search Console can illuminate part of it, not replace the rest.
Before approval, trace several statements in the automated report to specific rows or an aggregate, especially claims about “all”, “none” and “decline”. Add a condition that could change each interpretation. If the AI lacked the selected country or device, supply it and reassess the description. A stylistic edit cannot restore missing scope.
When combining Search Console and Analytics, retain separate definitions and retrieval conditions. The guide to GA4 thresholds, sampling and report scope (in Polish) (in Polish) helps check the second source. Matching column names in two files do not prove that they describe the same event.
Prepare a recheck package: original files, a filter record, transformation notes and the approved report version. Give them a shared identifier so a November summary is not compared with an October spreadsheet corrected later. A file hash in an AI report (in Polish) (in Polish) can identify exact bytes but cannot establish that data or interpretation is correct. Keep that limit in the evidence label too.
For each recommendation, record the finding that supports it. Repairing interrupted retrieval, checking URL mapping and changing an offer are different jobs for different people. If the material only supports the first, do not present the others as analytical findings. After a correction, return to the same question and check whether the earlier recommendation still has a basis. The decision owner should see what changed and what remains unknown.
The Aeneia Audit offering describes an operational review of reporting: sources, retrieval, calculations and responsibility for the approved account. It promises neither a Google ranking nor a substitute for specialist statistical analysis.
The public methodology was checked on 30 October 2026. All figures in the example are fictional; no private statistics from Aeneia or any other Search Console account were used.







