An AI answer cites a supplier page. Clicking the citation opens the expected website and highlights a sentence about delivery. A month later, the same link opens the same page, but the highlight has disappeared. A reviewer can still read the website. The original supporting passage is harder to establish.
This is a practical buying question for an AI visibility monitoring system. An archive needs to preserve enough information for another person to review the relationship between an answer, its exposed citation and the relevant source passage. A clickable address supplies part of that record. Text-location information supplies another part.
The difference becomes visible in links containing `#:~:text=`. These text directives can help a browser locate a passage inside a page. Their usefulness depends on the text, the document returned and the browser interaction. A monitoring export should give the reviewer a way to inspect those dependencies.
Separate the document address from the passage locator
Consider this invented citation address. The domain and sentence are teaching fixtures.
```text
https://example.com/supplier#:~:text=Delivery%20takes%20four%20weeks.
```
The portion after `#` is a fragment. RFC 3986 describes fragment handling as a client-side step separated from resource retrieval. The URL Standard represents the fragment as a component of the parsed URL. A reviewer can therefore preserve both the complete observed address and a separately derived document address used for retrieval or grouping. RFC 3986, section 3.5, URL Standard
This distinction gives a source record two useful questions. Which document did the request obtain? Which location did the citation ask the reader to inspect?
A document-level source count can group two locators under one page while retaining their original strings. The earlier guide to citation URL processing covers normalization decisions. The task here begins after preserving the address and asks whether the passage remains identifiable.
Read the text directive before trusting the highlight
The WICG draft describes text directives with a required starting text, an optional ending text and optional surrounding context. The text parameters are percent-decoded. With only starting text supplied, the described target is its first matching instance. Context terms help distinguish repeated passages. The document is a Draft Community Group Report. URL Fragment Text Directives
For the fictional delivery sentence, a simple directive and a contextual directive could look like this.
```text
#:~:text=Delivery%20takes%20four%20weeks.
#:~:text=Series%20production-,Delivery%20takes%20four%20weeks.
```
The second string adds preceding context. A reviewer should inspect the resulting passage in the actual destination and browser. Google's implementation article also describes navigation requirements, including a `noopener` context. Record how the link was opened when reproducing an observed highlight. Google web.dev implementation guide
A failed highlight leaves several possibilities to investigate. The destination text may have changed. The intended passage may occur in a different representation. Navigation conditions may affect the browser's handling. The source page and the visible result provide evidence for distinguishing these possibilities.
Repeated wording can point at different offers
Suppose a fictional supplier page contains two sections.
| Section | Sentence | Nearby qualification |
|---|---|---|
| Prototype programme | Delivery takes four weeks. | Applies after prototype approval |
| Series production | Delivery takes four weeks. | Applies to a defined repeat-order quantity |
The same short sentence appears twice. The commercial meaning depends on its surrounding offer.
A citation that exposes only the delivery sentence leaves the reviewer with two candidate passages. A location chosen by navigation may be convenient to display, while the reviewer still needs to establish which offer the answer describes. Reading the section and its qualification is part of checking the claim.
Now imagine three changes to the fictional page.
First, the publisher reverses the section order. Both sentences remain. A starting-text locator can lead the reader to a different offer.
Second, the publisher keeps the section order but changes one delivery sentence to six weeks. A stored four-week quote can still match the other section. The surviving match needs its own applicability review.
Third, the publisher leaves the delivery sentence intact and changes the eligible order quantity nearby. Text matching succeeds, while the condition underlying the buyer's decision has changed.
These cases call for three recorded findings. Preserve the observed locator, identify its candidate passage in the inspected representation, and review whether that passage supports the answer's actual claim. Each finding answers a different part of the investigation.
A quote selector provides another representation of the location
The W3C Web Annotation Data Model defines a `TextQuoteSelector` using selected text and optional preceding and following snippets. It specifies normalization before recording the text. Its treatment of multiple remaining matches preserves that ambiguity. An archive can use this model as a reference when designing passage records. Web Annotation Data Model, text quote selector
The following original JSON illustrates a buyer's proposed record for the fictional series-production passage.
```json
{
"type": "TextQuoteSelector",
"exact": "Delivery takes four weeks.",
"prefix": "Series production ",
"suffix": " Applies to repeat orders."
}
```
This is a proposed annotation for the example. Its relationship to a browser text directive should be documented by any system that converts between the two formats. The selected quote, its surrounding text and the representation from which they were taken need to remain associated.
The prefix in this record includes a trailing space. The suffix includes a leading space. Keeping those characters explicit makes the local exercise reproducible.
Run a small check against saved plain text
The fixture below checks literal, case-sensitive strings in a supplied plain-text representation. It reports every surviving range. It is deliberately small enough to inspect and adapt to a buyer's export acceptance exercise.
```python
def locate_quote(text, exact, prefix="", suffix=""):
if not exact:
raise ValueError("Provide a nonempty quote")
matches = []
start = 0
while True:
position = text.find(exact, start)
if position == -1:
break
end = position + len(exact)
before = text[:position]
after = text[end:]
if before.endswith(prefix) and after.startswith(suffix):
matches.append((position, end))
start = position + 1
return matches
quote = "Delivery takes four weeks."
prototype = "Prototype programme " + quote + " Applies after approval."
series = "Series production " + quote + " Applies to repeat orders."
saved_text = prototype + "\n" + series
assert len(locate_quote(saved_text, quote)) == 2
assert len(locate_quote(saved_text, quote, "Series production ")) == 1
assert len(locate_quote(saved_text, quote, suffix=" Applies to repeat orders.")) == 1
assert locate_quote(saved_text, "Delivery takes six weeks.") == []
print("Two bare matches; one match with the selected context")
```
Its output gives the reviewer a concrete result to compare with the saved representation. Two bare matches expose a repeated phrase. One surviving contextual match identifies a candidate for the next claim-support check.
The fixture uses Python string positions for this particular input. A browser's text matching, rendered-text construction and navigation behaviour need their own implementation checks. Record the extraction method when turning HTML or a browser view into the plain text supplied here. The existing citation offset exercise covers position units and bounds.
An acceptance rule can require a reviewer to resolve multiple matches before attaching a passage to a claim. That is a stated review convention. The programme's return value remains useful even when the convention changes, because it retains the candidate ranges.
Keep the representation beside the locator
A locator describes a selection within some content. Replaying it against today's website establishes what it selects today. A historical source review also needs the representation available at the relevant observation time.
The following fields are an editorial proposal for an acceptance record. A vendor can provide equivalent information using its own schema.
| Record element | Review question it answers |
|---|---|
| Original exposed citation URL | What address and locator did the interface provide? |
| Retrieved document URL | Which destination was inspected? |
| Observation and retrieval times | When were the answer and source each observed? |
| Authorized saved representation and its digest | Which retained bytes or text are being reviewed? |
| Extracted passage and extraction method | What material did the reviewer actually read? |
| Candidate matches and support decision | Was the passage identifiable, and did it support the claim? |
A digest makes a retained object's identity checkable. The record still needs a readable source, its capture context and a reasoned support decision.
For a changing supplier page, keep a historical review and a current review as dated records. Reassessing present delivery conditions is a legitimate buyer task. Reconstructing an earlier recommendation is another legitimate task. An archive becomes easier to use when the interface makes the chosen time basis explicit.
Include the locator in a software purchase exercise
Start with one authorized answer that exposes a source link containing a passage locator. Save the complete answer and the full exposed URL. Record the interface in which the link appeared and any source-card relationship the reviewer can actually observe.
Open that link through the normal user path. Record the destination, browser version, opening method and visible result. Inspect the surrounding passage and write a short decision about its support for the answer's claim.
Export the record through the product's normal account workflow. Compare the original URL with the exported value, including the fragment. Check whether a downstream spreadsheet or URL-cleaning step preserves it. Then repeat the passage review using the retained representation available to the reviewer.
Add the fictional repeated-sentence fixture when testing ambiguity handling. Ask the operator to show both candidate passages and explain the selection decision. Keep a changed-page example so that current matching and historical reconstruction can be demonstrated separately.
For Semvi by Xindar, the product team has confirmed complete-answer and citation-URL exports. This exercise provides additional buyer acceptance criteria for the exposed locator, retained representation and passage relationship. A live account demonstration establishes which of those criteria the selected workflow and export satisfy.
FAQ
Does a missing highlight mean the citation URL is broken?
Record page retrieval and passage-location results separately. A page can open while the intended text is unavailable or the browser interaction produces a different result. Inspect the current destination and compare it with any authorized retained representation.
Should the export keep everything after the hash sign?
Preserve the authorized observed citation string before deriving a grouping key. The fragment can carry the location that the reader was directed to inspect. Store a derived document address alongside it when the reporting task needs one.
Can a matching sentence settle whether the source supports the answer?
Read the sentence's section, conditions and relevant document version. The fictional four-week delivery sentence appears in two offers with different qualifications. The support decision depends on the offer described by the answer.
What should happen when the same quote matches twice?
Retain both candidates and apply the stated review rule. Context can narrow the selection. A reviewer should be able to see how the chosen passage was associated with the claim.
Sources and method
The technical references above were checked on November 16, 2026. The delivery examples, JSON record, literal matcher and acceptance procedure are original teaching fixtures. Two retained source-card URLs in a separate first-party research session contained text directives and motivated this topic. Those links led to the pages below. Their presence establishes the observed link form in that session. This article evaluates citation review and export usability. Geoptie source page, Enrich Labs source page
